thekb.eu — Transformation & Adoption

AI4IT vs AI4Business : le renversement, et ce qu'il fait à vos budgets 2027

Wed, 24 Jun 2026 00:00:00 GMT

In this opinion piece published on sfeir.com (June 24, 2026), **Didier Girard** (Managing Director of SFEIR) argues a thesis: the **AI4IT vs AI4Business reversal**. In 2024, the consensus saw **AI4Business** — AI poured into business processes (sales, support, finance) — as the great reservoir of productivity; **AI4IT** (AI for producing the information system) was seen as a topic for engineers. Two years later, *"the numbers have settled it, and the other way around."* **The AI4Business disappointment**: the 2025 MIT study ("95% of GenAI pilots with no ROI") is, by Girard's own admission — he disputes its methodology — questionable, but its **persistence** is the real signal of genuine dissatisfaction: many executives do not see the promised value in their processes. *"The symptom is true even when the figure is false."* The blockage is **organizational** (Mollick's Hayekian problem), not technical. **The AI4IT reversal** rests on quantified evidence: Salesforce (+151% Effective Output, migration 18× faster, −5% incidents), Intercom (3× R&D productivity, −50% cost/PR), Raiffeisen Bank Ukraine (−8% headcount but 7 new products, −70% blocking incidents), AWS (Bedrock redeveloped by 6 people in 72 days), Atlassian (+19 to +87% PRs), DORA × Google Cloud (39% ROI, 8-month payback). **Why?** Code **verifies itself** (compilation, tests, CI); business processes do not. *"We equip those who already know how to equip themselves."* **The 2027 budget consequence** comes down to three breaks. (1) **CapEx→OpEx**: the token becomes a variable OpEx charge — Arthur Mensch (Mistral) puts it at ~10% of payroll budget spent on tokens among advanced adopters. (2) **Token pricing, a double trap**: at a given capacity, inference has been divided by ~280 in two years, but the ceiling is rising (Fable 5 at $10/$50 = 2× Opus 4.8), while open models (GLM-5.2) and desktop inference push costs down; the Jevons paradox drives consumption up faster than the price falls. (3) **AI FinOps**: think in terms of **cost per outcome**, allocate by rules, turn token→outcome attribution into an asset. Four COMEX recommendations: fund AI4IT first (payback < 1 year), budget for the J-curve, put token FinOps in place before drift sets in, redefine headcount accounting (humans + agents). Conclusion: *"the next budget battle will not be about the price of the token, but about the cost per outcome."*

Comment l'IA agentique bouscule les Grands Groupes ? Partie 2/2 #DevSummit

Thu, 18 Jun 2026 00:00:00 GMT

The second part of an episode of the « À la French » podcast recorded at DevSummit, this interview brings together Mathieu Grymonprez, Global CDO of the Adeo group (Leroy Merlin, Obramat, Weldom), and hosts Jean-Baptiste Kempf (creator of VLC), Steeve Morin and Mehdi Medjaoui. Mathieu, 26 years with the company and 8 years as CDO, traces a career path from network-security engineer (first Check Point firewall) to leader of « Digital Tech and Data »: after urgently resolving an Oracle database crash in Brazil (2012) and then overhauling the local IS during six years as an expatriate, he rationalized the group's 24 IS / sites / PIM into digital platforms (customer & commerce, supply chain, retail, corporate) supported by a tech radar, documented APIs and microservices (which became « big products »). His thesis: **every transformation is won on two simultaneous fronts, culture and structure**, and the digital transformation playbook (waterfall → agile, product, more make than buy) is being replayed with AI. On the culture side: reconfigure to embrace the technology, keep critical judgment and above all **accountability** — responsibility remains human, « it's not the agent's fault ». On the structure side: close the documentation debt, manage agents' rights and permissions. Remembering the failure of the « Retail Apocalypse » (Amazon, e-commerce negotiated too late), the watchword is « we won't get caught out again »: take AI seriously, but with the same values (pragmatism, customer service, leading brand). If ChatGPT builds a better basket than the in-house app, « that's my problem ». At the board, Mathieu never talks technology but customer experience and ROI; he doesn't even ask for an AI budget, funding the new work through the gains (compressing JIRA tickets), in a logic of reuse serving the in-store salesperson. He doesn't anticipate the end of developers but an avalanche of requests (P10 projects become P2). On costs, he is confident: token FinOps will follow the path of cloud FinOps, driven by inference chips (TPUs) and open-source models catching up (Gemma 4 on a laptop). But model variation is a real production problem (retesting, requantization, silent downgrades), and Google has a « production awareness » that OpenAI or Anthropic don't yet have. His biggest concern: **enterprise intelligence lock-in** (agentic harness, « adeo.md »), hence the attention paid to standard Kubernetes, API portability and memory. He points to the missing open-source building block — agent orchestration (registry, lifecycle, permissions, skills) — and company memory (« when it's not logical, it's historical »). Final advice: transformation is bespoke; understand the technology mainly to avoid getting « fleeced » by pickaxe sellers.

AI made your engineers fast. Too fast to leave room for the rest of the org to think.

Wed, 17 Jun 2026 00:00:00 GMT

In this LinkedIn post, Fred Plais (co-founder and CEO of Archie, former CEO of Platform.sh) endorses and extends an observation about AI's real effect on tech organizations: by making engineers extremely fast, AI eliminated the thinking time the rest of the company used to have. The **bottleneck has not disappeared, it has moved upstream**, into an area nobody is watching. The reasoning starts from a historical observation. For years, **execution was the slow part** of the work: building something took long enough to leave room for thinking. Product people could read analyst reports, talk to customers, study the competition, and shape a genuine point of view **before** much code was written. That margin has nearly disappeared. The difficulty has therefore shifted: it now consists of having the right vision and making the right choices in a fraction of the time once available. From this shift, **two new rare profiles** emerge. The first knows how to articulate a clear vision, **precise enough for an agent to execute without derailing**: an agent builds exactly what it is asked for, and nothing more — "knowing what to ask for is the hard part." The second knows how to **properly orchestrate agents**: they know the agents' failure modes, know how to chain them, and can catch an error before it propagates. This second profile is more recent and still rare. Plais highlights the market's disconnect: many teams keep **hiring for "code output,"** even though that is precisely the resource that has stopped being scarce. The post's punchline is a moral: **thinking clearly has always been the job**; speed did not invent anything, it simply made it impossible to fake. Fred Plais adds his own comment: people keep asking him what AI changes for development, and his answer is "nothing" — but you can no longer fake it. He closes with a **driving metaphor**: driving at 200 km/h instead of 100 requires good brakes to avoid an accident and a perfect map to know where you're going. In other words, AI's acceleration of execution does not lighten the demands on judgment: it hardens them, shifting value toward clarity of vision (the map) and mastery of the guardrails (the brakes).

How Cornell Recovered $100,000 in Unidentified Payments With AI

Mon, 15 Jun 2026 00:00:00 GMT

The Cornell AI Innovation Hub recounts (June 15, 2026) how a two-semester collaboration made it possible to **recover $100,000** in unidentified payments using AI. The problem: every year, Cornell receives hundreds of wire transfers and ACH payments without enough information to route them (no invoice number, vague vendor name). The funds accumulate in a suspense account — active backlog ~**$1M**, historical peak **$4M** — and **New York State law mandates escheatment** if they are not resolved in time. Two treasury staff members were spending up to **half a day** a day on this. The project's structure illustrates Ethan Mollick's **Leader-Lab-Crowd** framework. The **Lab** is the **AI Hub** (Pete Stergion and Phil Williammee, co-tech leads, plus a cohort of students). The **Crowd** is **Treasury** (Cheryl Barnes, Marie Graves, Kevin Mooney, Debra Federation), holder of the business knowledge and the data — Kevin provides **3 years of Oracle GL history (10,000+ records)**. The student analysis surfaces the key insight: **99%** of payments carry a vendor name, versus **less than 4%** an invoice number. The build follows a **"context first, then plan, then build"** discipline: via **Claude Code Plan Mode**, the team loads all the context (notes, manual process, prototypes, sanitized data); Claude Code **proposes an architecture to validate before writing any code**. A semester of notes becomes a **working tool in a single session**. The **Python pipeline** (exposed as a *skill* `/treasury`) chains three steps: **fuzzy matching** against the GL (filtering out noise words like Inc/LLC/Corp), **vendor lookup** via **Gemini Enterprise Web Search**, then **Claude synthesis** producing, for each payment, a likely department, a **confidence level**, and a contact. Output: an Excel file sorted by confidence, in a few minutes — all within the governed **Cornell AI Gateway** (PII stripped, no external model training). The **backtest** (9,131 resolved payments) shows **97% → 100%** accuracy for recurring vendors with the full AI chain, and **76% → 100%** for unknown vendors. Documented limitation: vendors billing multiple departments. Operational result: 23 departments contacted, 7 responses, **5 payments = $100,000** confirmed. Beyond the figure, the case is a **counter-example** to the narrative that "AI doesn't create business value": it does here, because a **Lab**, an expert **Crowd**, and **real groundwork** came together. And the $100,000 plays the role of the **visible reward** Mollick prizes — the tangible proof that legitimizes and spreads adoption, by removing the drudgery rather than the jobs. *"The $100,000 is a start."*

The AI-native SDLC is paying off: 19% more PRs and 2–3 hours saved per developer per week

Sun, 31 May 2026 00:00:00 GMT

On its Inside Atlassian blog, Atlassian publishes a data study co-authored by two data scientists (Robbie Geoghegan, Fan Jiang) measuring the actual return of an **AI-native SDLC** powered by its **Rovo Dev** agent. The stakes are framed from the outset around the "productivity paradox" formulated by Robert Solow in 1987 ("you can see the computer age everywhere but in the productivity statistics"): AI is massively adopted — 93% of developers use AI tools, nearly 30% of code is written by AI — but its impact remains unclear as long as it is measured in **usage** (tokens) rather than **impact**. The results, drawn from a quasi-experiment across 3,400 repositories from 2,500 customers (propensity-score matching), are quantified and segmented. Repositories adopting Rovo Dev merge **19% more pull requests per month** than non-adopters. The gain rises to **37-51%** on low- or medium-activity repositories, and **doubles to 59-87%** when **3 to 5 members** of the team adopt the tool: collective adoption clearly outperforms individual adoption. On the efficiency side, a survey of more than 6,200 developers (estimates taken at the 20th percentile, hence conservative) establishes a gain of **2-3 hours per week** on coding and review tasks, or about 10% of the 24 hours they involve — that is, 20-30 hours per week reinvested for a team of ten. The article proposes a **five-stage AI-native SDLC** in which the agent supports the human: Plan (proposed breakdowns and estimates), Orchestrate (human/agent coordination), Code (autonomous agents on well-scoped work, PRs ready for review), Review (review against team standards before the human) and Operate (always-on incident copilots). It pairs this with a **four-dimension measurement framework**: Speed (PR throughput), Efficiency (time saved), Quality (change failure rate) and Satisfaction (developer satisfaction) — so as not to reduce value to velocity alone. Two points reinforce the argument. First, the role of **context**: thanks to Atlassian's Teamwork Graph, context-rich AI delivers results that are 44% more accurate while consuming 48% fewer tokens. Second, the **operational recommendation**: start with a team (not an individual), choose a repository with 3-5 engineers who are actual users, and measure throughput and time savings 2-3 months after deployment, once the novelty effect has worn off. The underlying message: the value of AI is real but conditioned on rigorous impact measurement and team-level adoption.

After Automation

Thu, 21 May 2026 00:00:00 GMT

Pivotal essay by **Dan Shipper** (CEO Every) published on **May 21, 2026** on every.to, *"After Automation"* — an argued response to the thesis of AI-driven collapse of knowledge work. **Pivot thesis**: AI progress creates **more work for humans, not less**. Looping mechanics (***"the commodification cycle"***): (1) AI commoditizes yesterday's human skill; (2) that cheap skill is widely adopted → abundance; (3) abundance produces *sameness* (the *"slop"*); (4) humans demand difference → renewed demand for experts; (5) experts use AI to address today's problems → loop. **Canonical quote**: ***"There's more work to do than ever"***; ***"AI commoditizes the residue of human expertise, creating demand for what's different"***. **Central conceptual framework — Frame vs. Framer**: benchmarks measure performance ***"within frames"*** (specific problem framings); once saturated, *changing the frame resets the counter* — models **escalate within frames but do not replace the framers**. Pivot formula: ***"the frame is not the framer"***. Even at AGI, humans must **specify goals and interpret results** — *"the frame problem regenerates one level up"*. **The "Human Sandwich"**: Human sets frame → AI executes → Human judges and extends. **Two modes of working with agents**: (a) ***agent employees*** — asynchronous delegation (coworker / embedded — Claudie, Andy, Viktor, Fin); (b) ***human-AI collaboration*** synchronous (Claude Code and equivalents). **Every data**: 95% of CEO emails processed by AI; **Fin (Intercom) resolves 65% of support conversations**. **The Zeno's paradox of AI**: AI continuously closes the gap, but humans remain "the turtle ahead" because they are ***"alive to a specific moment"*** — *"running wants, running concerns"* — while models operate on historical training data. **Detailed benchmarks**: **GPT-5.5 = 62/100 on Senior Engineer codebase rewrite** (vs human 80-90s); **GDPval**: 40-49% of expert human level, **but with extensive human framing**. **OpenClaw 44,469 PRs** in May 2026 (vs Kubernetes 5,200 in 2022) — proof that agentic work creates *"more work"*, not *"less human work"*. **AGI implications**: even at AGI, the **human framer** remains structurally ahead — addressing *"current, situated"* problems while the model operates on *"historical training data"*. **Anti-tipping-point pivot conclusion**: this is not a tipping-point event, it is ***a persistent pattern*** that defines the future of work. **Major relevance**: an explicit counter-narrative to *Amodei white-collar bloodbath* / *Sun permanent underclass* / *Anthropic Economic Index* — Shipper, **CEO of a company that lives with agents daily**, offers the theoretical framework that reconciles the two empirical observations (AI does more + humans remain indispensable). Strong convergence with **Ng "No AI jobpocalypse"** (2026-05-08), **Mollick × roon ASI / FDE** (2026-05-10), **Tatsyi/Raiffeisen "AI made engineers different"** (2026-05-05), **Curran/Intercom 3× R&D** (2026-04-16) — all describing humans as *redeployed toward framing* rather than *replaced*. Productive tension with **Sun NYT permanent underclass** (2026-04-30), **Wallace-Wells AI populism** (2026-05-08), **Osmani Cognitive Surrender** (2026-05-05 — the human framer must remain active). To be leveraged for COMEX / DG / boards: strategic vocabulary for 2026 — *"frame vs framer"* becomes the canonical grid for AI governance.

AI-assisted engineers are burning out, is this fine?

Tue, 19 May 2026 00:00:00 GMT

Pivot article **Ivan Chepurin & Travis Turner** (Evil Martians Chronicles, **May 19, 2026**) — ***« AI-assisted engineers are burning out, is this fine? »*** — **structured diagnosis of burnout among AI-assisted developers** and a **5-axis intervention toolkit**. **Pivot thesis**: AI-accelerated productivity hides a **hidden cost — developer exhaustion**. *« Higher productivity doesn't translate to sustainable work practices or job satisfaction. »* Shunryu Suzuki epigraph on mental agitation. **TL;DR — 3 essential remedies**: (1) restore enjoyment of the process, (2) rebuild accomplishment / ownership / pride, (3) remove the pressure of continuous productivity maximization. **Central narrative frame — Ben vs Alice**: Ben (traditional coding) = 4 h of steady work, distributed cognitive load, satisfaction at completion; Alice (AI-assisted) = 2 h of cognitively high-intensity work, continuous task-switching, **no satisfaction** + fills the freed-up time with more tasks → **exponential escalation of load** despite accelerated output. **Canonical formula**: ***« We compensate for a lack of satisfaction with work quantity. »*** **Structural disruption of the craft cycle**: (planning → crafting → result) compressed into (planning → result), removal of the meditative craft phase replaced by **cognitively demanding code review**. Direct convergence with **HBR study 2026** (cited): *« cognitive exhaustion from intensive oversight of AI agents is both real and significant »* + **UC Berkeley research 2026**: workers fill natural breaks with AI tasks. **Quiet career change** — pivot concept: developers hired to code now do **different work without a conscious career transition**. 4 possible paths: (1) find enjoyment in the new structure (prioritized), (2) ignore AI, (3) work without enjoyment (unsustainable), (4) change careers. **5 daily burnout factors identified**: (1) ***Losing context*** — the agent carries project understanding externally, cognitive-debt shift from code to people, loss of system intuition; (2) ***No time for passive thinking*** — *« The model fills the silence before your own thinking has a chance to connect dots »* (showers, walks eliminated as moments of unconscious problem-solving); (3) ***False expectations*** — initial speed = unrealistic baseline, subsequent slowdowns experienced as failure; (4) ***Review bottlenecks*** — *« the more code is generated, the more code needs to be reviewed »*, disproportionate cognitive load on seniors, diffusion of responsibility; (5) ***Endless possibilities*** — low prompting friction encourages constant pivots, absence of natural scoping. **5-intervention toolkit**: (a) **Acknowledge your wins** (win-log, team demos, hours tracker); (b) **Rethink AI workflow** (planning > review, **3-4 iterations max**, no parallel task-switching, separate AI-heavy tasks with breaks, decompose); (c) **Keep exercising your craft** (protected AI-free craft hours, *« ask » mode > generation mode*, agents off on passion projects); (d) **Discipline + work-life balance** (fixed hours, real breaks, daily intentions, stop when done); (e) **Find new areas of interest** (user research, soft skills, analytics, agent fine-tuning + guardrails, perf optimization). **Conclusion**: *« AI can be helpful. Problems appear only if you misuse it. »* Industry evolution = inevitable; individual well-being = controllable. Major convergence with **Osmani Cognitive Surrender** (2026-05-05), **Frizzo "Year With Claude Code"** (2026-05-05 — *« writing muscle atrophy »*, *« deep flow rare »*), **Bedard BCG/HBR Brain Fry** (2026-03-05 — 1,488 employees, peak of 3 tools, +39% errors, +39% intent to leave). Major relevance for **CTO / VP Engineering / IT HR** dealing with the retention of AI-augmented engineers in 2026.

AI/works™ by Thoughtworks — Thoughtworks' Agentic Development Platform / "We are doing it again for the AI era"

Tue, 12 May 2026 00:00:00 GMT

Thoughtworks launches **AI/works™**, its **agentic development platform** claimed to be *"the new standard for building and running industrial-grade systems in the AI era"*. The opening tagline — ***"We are doing it again for the AI era"*** — explicitly invokes Thoughtworks' heritage (XP, Continuous Delivery, microservices, refactoring) to sell the new platform. The **core thesis is economic**: *"The breakthrough is the economics. The old approach made you pay millions to build, run, then pay again to rebuild. AI/works™ ends that routine."* Four derived promises: continuous system updates, selective regeneration *"without the token blowout"*, *"your systems finally stop aging"*, and fast-tracked new products. The platform deploys **six capabilities** covering the entire SDLC: (1) **Reverse Engineering** ingests legacy code and produces validated as-is specs; (2) **Dynamic Spec Development** converts raw requirements into a unified *Super Spec* covering architecture, workflows, security, data, UX; (3) **Spec to Code** generates testable code from the Super Spec via *coordinated agents*; (4) **Developer Experience** standardizes AI-assisted development via *governed golden paths*, automated pipelines and a shared catalog; (5) **Control Plane** orchestrates and governs agents with *cost transparency*, *active guardrails* and *end-to-end lineage*; (6) **Runtime Ops** continuously monitors, detects change, updates the Super Spec and regenerates impacted code. The pivot concept is the **Super Spec**: a **dynamic, unified specification** serving as the source of truth, automatically updated in production and triggering regeneration of impacted code rather than patches. The **3-3-3 methodology** structures delivery: 3 days to align the product concept, 3 weeks for a prototype (desirability/viability/feasibility), 3 months for a production MVP. ***"Industrial-grade systems that grow up instead of grow old."*** **Constellation Research** recognizes AI/works™ *"for changing the economics of enterprise software delivery"* via a *"spec-driven, lifecycle"* approach. The **anti-positioning** is explicit: *"no consultant crowds"* (a direct jab at the large integrators), *"finance can open the bill without switching on emergency lighting"* (corporate self-deprecation), *"stands on an engineering foundation rather than enthusiasm"* (deliberate anti-hype). The **featured partners** span the entire agentic stack: AWS, GCP, Azure, Databricks, Snowflake (cloud/data), Claude, DeepSeek, Gemini, Grok, OpenAI (LLMs), NVIDIA, Groq (compute), CAST, Mechanical Orchard (legacy), Stripe, Spotify (likely reference customers). The **dual CTA** *Request a discovery call / Sign up for updates* betrays a **sales-led, high-ACV** model. Read within the 2025-2026 corpus, AI/works™ is the **productization** of the Thoughtworks doctrine carried intellectually by Kamelman (*Service-as-Software*, 2025-12), Fowler (*LLM Retreat*, 2026-02) and Böckeler (*Harness Engineering*, 2026-04). It is the **Anglo-Saxon commercial equivalent** of the Wescale doctrine *Usine Logicielle Augmentée* (2026-05-03), packaged as a platform.

You will know that the AI labs believe in ASI when [they dissolve their forward deployed engineering teams]

Sun, 10 May 2026 00:00:00 GMT

Ethan Mollick (Wharton) launches a consistency test aimed at frontier AI labs on LinkedIn: *"we'll know labs truly believe in ASI when they dissolve their Forward Deployed Engineering teams"*. As long as humans are needed to integrate AI into client organizations, white-collar jobs are not threatened in the short term. **The debate with roon** (OpenAI employee, influential voice of the *accel* circle on X): roon retorts that this is a **hayekian problem**. Reference to Hayek (*"The Use of Knowledge in Society"*, 1945): useful information within an organization is tacit, distributed, contextual. Central intelligence, even superintelligent, does not automatically resolve information flows. roon says he is more optimistic about employment than the average lab precisely for this reason. Mollick concedes the hayekian point then turns the argument around: if AI is not self-adopting, then the lab prediction that "most white-collar jobs will be replaced by 2035" is contradicted by their own FDE teams. roon sums up the apparent agreement: *"Gentle singularity"* — the transition will be slow and mediated, not a *fast takeoff*. Sam Altman will later reuse this term in his essay *The Gentle Singularity* (June 2025). **The consensus in the comments** (practitioners, consultants, researchers) converges on four points: 1. Technology is often the easy part. The real obstacle is internal politics, HR incentives, legacy systems, and above all the question "who is liable when it breaks?". 2. ASI can produce a perfect transformation plan and still get stuck on a VP who refuses to modify their Salesforce workflow. 3. Accenture (and consulting more broadly) survives because it sells **contractual liability**, not just competence. An AI cannot be sued. A firm can. 4. Most shared phrase: *"Curing cancer might be easier than replacing Accenture"* — a technical problem has clear success criteria; an organizational problem does not. **The structuring tension** that Mollick formalizes: **East Coast vs West Coast** (epistemic, not geographic). East = slow, fragmented transformation, constrained by the *jaggedness* of capabilities and social complexity. West = fast, massive automation as soon as capabilities are sufficient. **Conclusion**: labs sell ASI but hire consultants. This is either a logical contradiction, or — a more cynical view — a short-term revenue strategy that funds the long-term bet. Either way, their own FDE teams attest that AI is not (yet) self-adopting. The adoption bottleneck migrates from the technical to the organizational.

IA : et si les développeurs disparaissaient ? — Tech & Co Business, Le débat (BFM Business, 05/05)

Tue, 05 May 2026 00:00:00 GMT

On the set of *Tech & Co Business* (BFM Business, May 5, 2026, "The Debate" segment, 17 minutes), Rémi Jacquet (CEO of Cast Software France, founder in 2023 of a think tank of about a hundred CIOs partnered with Cigref/Epita) and Didier Girard (CTO and CEO of SFEIR, a French IT services company of 1,000 people) compare their readings on the future of the developer profession in the age of AI agents. Girard strikes from the outset: ***"writing code has become an anti-pattern"***. AI now produces code of higher quality than most engineers and is *"2 to 10 times more efficient"* — *"that's the reality, you have to accept it"*. But the profession is not disappearing: *"we delegate analysis and reasoning to AI, the decision remains human."* Jacquet adds: the developer becomes a ***conductor / agent manager / arbiter***, talking directly with the business and taking full responsibility for an application. The 14-day Scrum sprints give way to one-hour to half-day ***bolts***. Amazon's **Pizza Team** (8-10 people) no longer holds: *"if we put 10 people together to empty a dishwasher, we won't go any faster"* — teams need to be **segmented**. SFEIR claims *"1,000 people for a production capacity of 10,000"*. A new role is emerging — the ***product engineer*** —, measured not in lines of code but in value created. On the risk side, Jacquet raises a counter-intuitive warning: ***"the stronger AI becomes, the more we let our guard down — the more risks there are."*** Cast positions itself on ***harness engineering***: structured **deterministic** analysis to channel **probabilistic** AI and control the code produced (architectural consistency, guardrails). Announced historical pivot: *prompt engineering* (2024) → *context engineering* (2025) → ***harness engineering*** (2026), aligned with Sylvain Duranton's (BCG X) op-ed published the same day in *Les Échos* (*"an agent = an LLM + harnesses"*) — explicitly *"French technology"*. On the HR side, **the token becomes the fuel of value**: NVIDIA allegedly pays bonuses in tokens, *"a developer who doesn't consume tokens is like a taxi driver without gas — is he really creating value?"* The lifespan of a skill drops from **10 years to 1 year**, skills-based job descriptions falter, HR departments become pivotal to the transformation, and a new think-tank group is working on a *"complete overhaul of the SDLC"*. Jacquet's provocative thesis: *"it might be easier to create a new IT department that is already AI-agent-based than to transform an existing one."* Finally, to juniors: *"there are positions to be taken — provided you become a conductor and master the fundamentals of software architecture."* Girard concludes: ***"code is the score — you need to master the symphony."***

A Year With Claude Code: My Output Doubled. My Attention Span Didn't.

Tue, 05 May 2026 00:00:00 GMT

Alexandre Frizzo published a *one-year retrospective* op-ed on LinkedIn Pulse on May 5, 2026, on his daily use of Claude Code. The title condenses the thesis: ***"My output doubled. My attention span didn't."*** A nuanced assessment rare in the 2026 corpus — productivity multiplied by **3-5×**, but with **hidden cognitive costs** acknowledged. **The job has changed shape.** Frizzo no longer writes code; he **makes decisions about code generated by agents**. *"Write hard parts, review easy parts"* becomes multi-codebase supervision. The bottleneck has shifted: protecting the *deep work state* has become *irrelevant* now that code generation is fast; ***"the new bottleneck is supervision"*** — reading agent output, deciding correctness, integrating, catching subtleties. *"Quality control at scale now requires entirely different defenses."* **Acknowledged gains**: 3-5× productivity multiplier on a typical day, previously infeasible projects now accessible (yak-shaving, boilerplate, *long-tail*), near-zero cost of experimentation. **Acknowledged losses**: *writing muscle* atrophied (*"manual coding feels effortful now"*), deep flow state rare due to constant context-switching, diminished ownership satisfaction (***"the code is good, but isn't quite mine"***). **Statistics cited**: median 3-4h effective coding over an 8h day, **23 min of context recovery per interruption** (Gloria Mark), 15-25 min to enter flow, 500% productivity in flow (McKinsey senior executives study). **Unresolved tensions** that Frizzo raises without settling: (1) **FOMO** — *"every hour I'm not at the keyboard is an hour an agent could be earning for me"* — the psychological pressure of 24/7 agents; (2) **Review quality** — reviewing at 3-5× the volume risks *skimming*, quality practices assumed a human pace; (3) **Skill atrophy** — *"does the writing muscle still matter, or is it becoming commoditized?"* **Exemplary epistemic stance**: Frizzo simultaneously rejects the two available narratives — *"AI is bad"* and *uncritical enthusiasm*. He holds an honest third position, real gains + real costs, **unresolved tensions** rather than conclusions. **Connection to the watch dossier**: a salutary complement to **Cherny's *"coding is solved"*** (2026-05) — same intensive daily use, but Frizzo sits on the **committed median** (3-5×) whereas Cherny and the Curran/Intercom top 5% occupy the **elite tail** (10×+). The two do not contradict each other: they measure **two different points** on the productivity distribution. Numerical convergence with the median: **Wescale** (2026-05-03), **Curran/Intercom** average (2026-04-16), **DORA Report 2025**, **Stanford Denisov-Blanch** (2025-11-23). Cognitive convergence with **BCG Brain Fry** (Bedard et al., 2026-03-05), **Anthropic study junior engineers deskilling** (cited by Sun, NYT, 2026-04-30), **Karpathy's *outsource thinking but not understanding*** (2026-04-29), **Soto's *Developer Taste*** (2026-04). To be used as a balanced practitioner testimony for executive committees, manager awareness, and the ethical debate on the developer's transformation into a supervisor.

PROJ-AI — pour que vos projets ne s'arrêtent plus au livrable (Un repo, un agent, un IDE — pourquoi PROJ-AI ?)

Tue, 05 May 2026 00:00:00 GMT

Antoine HABERT (WEnvision, a French agentic AI consulting firm) published a methodology article on May 5, 2026 that formalizes **PROJ-AI**, a lightweight methodological layer so that collective projects become **transferable rather than dying with their deliverable**. The diagnosis: projects stop at their deliverable, leave no organizational memory, critical decisions become untraceable, onboarding a new member takes three weeks. Central aphorism: *"The project is not a byproduct of the deliverable. The project IS the deliverable."* The methodology rests on a **triad**: (1) a **version-controlled git repository** as the single source of truth; (2) an **AI agent** (Claude Code, Cursor or equivalent) that reads the doctrine at every session; (3) a **markdown doctrine** specifying decision protocols and agent behaviors. The repo is organized into **six zones**: `DOCS/` (raw inputs), `IDEAS/` (hypotheses), `DR/` (Decision Records scored across 7 dimensions), `OUT/` (deliverables), `DOCTRINE/` (governance), `AGENT/` (slash-commands and session traces). The operational **DPEV** cycle — *Decide → Promise → Execute → Verify* — articulates four traceable steps that turn vague ideas into defensible deliverables. A **dual interface** combines PROJ-AI Studio (Cockpit/Guide/Weekly/Monthly views for the business side) and CLI/IDE with slash-commands `/dr-create`, `/livrable-update` (for tech), on top of the same single repo. Five **agent directives** structure behavior: ingest the doctrine before answering, explicitly cite internal sources, propose Decision Records for emerging choices, never override doctrine, produce end-of-session summaries. Across **three active engagements** (31 decisions recorded), HABERT documents: new-member onboarding **3 weeks → 2 days**, structural decisions tracked **30% → 100%**, architecture doc compilation **6 weeks → continuous**, PM archaeology time **30% → negligible**. The shared **proj-ai-commons** library capitalizes on anonymized patterns (DR templates, slash-commands, doctrine fragments) and lets a new project bootstrap in **30 minutes instead of a week**. Two honest *caveats*: (1) ***"technology 20%, team discipline 80%"*** — an explicit rejection of solutionism; (2) ***"agent-agnostic"*** — markdown supports multiple LLMs. Status: Studio in internal preview, methodology delivered as a guided engagement by the firm (not self-service). This piece, converging with **Wescale Usine Logicielle Augmentée** (2026-05-03) and HABERT's previous fiche (2025-10-29), consolidates a French doctrine of AI industrialization. PROJ-AI transposes to project/consulting management the patterns of *Skills* (Vincent), *Plugin Marketplace* (Curran/Intercom), *AGENTS.md* (Anthropic, Vercel, Osmani) and *Decision Records* (Wescale). To be leveraged as an operational framework for consulting firms, CIOs, and project teams on 6-18 month engagements.

AI didn't make our engineers just faster. It made them different.

Tue, 05 May 2026 00:00:00 GMT

Hryhorii Tatsyi, CTO of **Raiffeisen Bank Ukraine** (~900 IT engineers), published a **12-month** longitudinal account (May 2025 → April 2026) of his organization's AI transformation on Medium in May 2026. The title crystallizes the thesis: ***"AI didn't make our engineers just faster. It made them different."*** This is one of the **rare quantified organizational case studies from a regulated European bank** available in 2026. **Core data**: IT headcount contracted by **75 people (−8%, including 64 engineers)** — yet **more code shipped, fewer incidents, improved security**. AI adoption rose from **62% to 83%**; **68% of engineers receive ≥50% of their code via AI assistance**; new-engineer onboarding **60-90 days → ~40 days** (consistent with Anthropic data of 82→40 days, an **independent convergence**). **Three emerging archetypes**: (1) **Copilot-only**: +10-25% on PRs, stable scope; (2) **Multi-tool**: story points **×1.5-3**, cross-repo scope **+50-80%**; (3) **Claude on corporate stack**: code volume **×4.5**, radically expanded scope. Counterintuitive insight: ***"AI lifts underperformers to baseline"*** rather than mainly accelerating top performers — the distribution **tightens from the bottom**. Senior architects **return to active development** after years away from it. **Seven new AI products** (that did not exist before): Service Knowledge Hub (57 microservices, 83 releases/month), Mobile Android workflow CI, AI Agent Portal (2,085 users / 649 MAU in 87 days, MCP generation via OpenAPI), Shift-left Security Plugin (−82% exposed secrets), DevPortal Backstage + Kubernetes diagnostics agents (−68% critical incident resolution time), DRAIF MCP text-to-SQL Data Lake with 10,000 tables (embedding fine-tuned ×2 OpenAI), Call Evaluation (>97% accuracy, **voted best product in the Raiffeisen Bank International group**). Stability: **blocking incidents −70%, critical resolution −68%, high-severity security alerts resolved +155%**. **Strategic pivot thesis**: ***"AI expanded our production possibility frontier, and we deliberately allocated the freed capacity"*** — toward features, stability, and technical-debt repayment. **Reframed evaluation question**: not *"by how much % did existing KPIs increase"* but ***"what did your engineers build that didn't exist before"***. **Tie-in to the watch corpus**: numeric convergence around the committed median with Frizzo (2026-05-05), Wescale (2026-05-03), Curran/Intercom (2026-04-16), DORA 2025, Stanford Denisov-Blanch (2025-11-23). Independent convergence on ~40-day onboarding with Anthropic. Productive tension with Cherny / Curran top 5% / Karpathy (elite 10×+ tail): the distribution **tightens from the bottom** AND **widens at the top** — both readings coexist. Cross-cutting convergence on "the job is changing shape" with Frizzo, Karpathy, Mornati, Habert. To be used for banking/regulated-sector executive committees, transformation sponsors, and the productivity-distribution-equity debate.

Slider Augmented Dev — La chaîne de production augmentée : comprendre la révolution de la chaîne de production logicielle à l'ère de l'IA

Sun, 03 May 2026 00:00:00 GMT

The French consulting firm Wescale presents its **Augmented Software Factory** doctrine: a software value chain entirely orchestrated by specialized AI agents, structured across six production lines. **Intention** ingests business needs into PRDs and ADRs; **Plan** derives User Stories and parallelizable tasks; **Sign-Off** is the sole human gate before production, where architects and Product Managers validate across six dimensions; **Production** runs 24/7/365 in parallel with integrated Shift-Left; **Verification** is entrusted to an independent audit agent comparing spec vs code (*"no self-certification"*); **Deployment** closes the loop with automated DevOps and AI monitoring. Key principle: **humans intervene at only two moments**, everything else is automated, traced, auditable. Wescale claims the **return of the predictable V-cycle** against Scrum agility, which it argues failed to hold scope, cost and timeline. Figures cited: **3-5x faster** than a traditional team, continuous **24/7**, **100% of code auditable**. Maximum ROI centers on legacy modernization and replacement of costly SaaS. Five business models are compared; **outcome-based** (payment on KPIs achieved) is favored as aligned with value. The *Myths vs Realities* slide condenses the anti-hype thesis: not 10x but **realistic 3-4x**; not the disappearance of developers but an evolution toward ***Strategic Judge*** and ***Agent Manager***; not a simple prompt library but a **Gov+Prod+Audit platform**; not a replacement for offshoring but its next step; and above all *"very few organizations will get there"* because **injected governance is a real barrier to entry**. Five mindset shifts structure the transition: code producer → strategic judge, manual writing → plan validation, deterministic → probabilistic critique, solo developer → agent manager, accepted technical debt → continuous Shift-Left quality. Six new key skills: judgment & ethics, orchestration & prompting, architecture & requirements, data-driven steering (DORA), Shift-Left quality control, resilience & adaptability. **Velocity is replaced by Predictability**. Four structuring pieces of advice close the deck: (1) tooling strategy is not enough, method comes first; (2) **injected governance is the central innovation** — *"a near-military layer"* — *"you don't stop AI from misbehaving by hoping it behaves well"*; (3) control compute costs (Drift, Token Burning); (4) separate immutable principles from adaptable implementation. Four risks to anticipate: security (vulnerable code), compliance (GPL, GDPR), governance (invisible technical debt), operational (Drift, Token Burning). Doctrine proven internally: *"What we learned by building Solario on Solario."*

« On est dans une boîte de Petri » : la Silicon Valley, ce pays où les agents IA sont déjà des collègues

Wed, 22 Apr 2026 00:00:00 GMT

On April 22, 2026, Florian Dèbes published a report in Les Echos from San Francisco on the integration of AI agents as full-fledged colleagues at Silicon Valley start-ups. The sentence that structures the entire article comes from Aaron Levie, CEO of Box, quoted by the New York Times: *"Silicon Valley right now is a real petri dish."* The report first documents the boost. Justin Bangay (Airbyte salesperson) has Claude prepare every client meeting from previous recordings and the web: *"It takes a minute, I save almost half an hour."* An investment fund partner has Claude scrape LinkedIn and ZoomInfo before he wakes up to deliver a daily sales brief. Sarah Allali (Lobby) prepares her fundraising through Claude, which lists investors and shared LinkedIn contacts. Logistical emails are delegated to agents — some even sign their replies to make this explicit. But Allali immediately points out the blind spot: *"Humans have an ego. Nobody wants to know they're not important enough for someone to bother writing to them."* On the engineering side, Jérémy Chapeau (SubImage) reports having shipped five major features in one week — *"Without AI I would have shipped only one"*. He built his own agent named **Jarvis** (an Iron Man reference ubiquitous in the Valley) that orchestrates action plans and responds to alerts from another agent monitoring customer behavior. Patrick Joubert (Rippletide) practices maximal delegation, **5 parallel agent tabs**, and formulates the central aphorism: *"The limiting factor is human cognition."* Then comes the flip side. Babak Hodjat (Cognizant) notes that AI causes fatigue *"when you delegate too much, the result is mediocre, and you have to redo everything"*. The article relays a viral post by Siddhant Khare (Germany): *"You use AI to be more productive. So why are you so tired?"* His economic thesis: *"AI reduces production costs, but raises the costs tied to coordination, verification, and decision-making. And those costs fall entirely on humans."* A BCG/Harvard Business Review study (Julie Bedard, March 2026) puts at **14% the share of employees overwhelmed by the pace imposed**, with cases of **"brain fry"** (cognitive overheating, headaches, slowed decision-making). The **token-max** — an internal ranking that rewards the heaviest AI users — fuels this exhaustion. The article closes on an anxiety shared by the builders themselves: *"Those who adopt it, or even build it, wonder whether they're digging their own grave."* Eric Pantera (Swile, Montpellier), however, notes that the SF/Europe divide has largely disappeared for those willing to engage: *"The gaps with our friends at Meta aren't significant."* A pivotal 2026 piece on everyday life with AI agents, which simultaneously records the productive success and the first wave of cognitive exhaustion among early adopters.

The AI-native interview

Wed, 22 Apr 2026 00:00:00 GMT

Sierra, a conversational-agent startup co-founded by Bret Taylor, has redesigned its engineering interview process to reflect how the job has changed in the age of coding agents (Codex, Claude Code). The authors — Vijay Iyengar, Arya Asemanfar, and Angie Wang — argue that the engineer's role is shifting from "building the machine" to "designing and refining the machine," by analogy with how engineers stopped worrying about the compiler's translation from code to machine instructions. Now that a single engineer can build across the entire stack, value comes from combining technical capability, product thinking, and business context. The starting observation: the legacy process (two coding interviews, algorithms, system design, culture fit) mostly captured mechanics — typing syntax, recalling algorithmic details, assembling frameworks. That signal grew increasingly dissonant with the day-to-day reality of the work. Hiring managers compensated by falling back on referrals and prior experience. Three criteria guided the redesign: representativeness (reflects real work), high signal (clarity on where a candidate excels or needs support), and a positive experience. The centerpiece is a three-part "AI-native onsite." **Plan**: a working session where the candidate ideates a product in their domain, with interviewers asking questions to sharpen the idea. **Build**: 2 hours solo, using AI tools and frameworks of the candidate's choosing, with full freedom to pivot. **Review**: a demo, a discussion of product choices, a code review (data model, abstractions, extensibility), and a conversation about the path to production and how AI was used. Candidates are allowed to cut scope and skip boilerplate, per Paul Buchheit's formula: "if it's great, it doesn't have to be good." The rest of the process followed suit. The coding phone screen (no AI, in an online editor) is replaced by a system design interview — since vibe coding is easy, the real challenge is scalable production deployment. A "debugging interview" is being piloted to capture 1→N work in existing codebases: the candidate reviews a cross-cutting PR with agents. Lessons learned: hiring is for strengths, not the absence of weaknesses; debriefs have shifted from "should we hire this person?" to "where will this person excel?". Candidates report more engaging interviews — one built an AI flow game, a backend engineer drove their demo through an agent and a markdown file. Challenges (standardization, calibration) are mitigated by criteria that are agnostic to the product built and by paired interviewers. The format also applies to infra, where engineers now build full-stack and integrate vertically with the product.

The ROI of AI-assisted Software Development

Tue, 21 Apr 2026 00:00:00 GMT

Joint **DORA × delta** report (Google Cloud Professional Services), 60 pages, version **v. 2026.1** (citations retrieved February 2026, PDF created April 21, 2026), licensed **CC BY-NC-SA 4.0** — the first official **DORA ROI** framework dedicated to AI in the SDLC, with an **interactive calculator** at dora.dev/ai/roi/calculator. Pivot thesis: ***"AI is an amplifier"*** — AI **amplifies** the strengths of high-performing organizations and the dysfunctions of struggling ones simultaneously; it does not create performance, it **multiplies it where it already exists**. New central concept: the ***J-Curve of AI value realization*** — every AI adoption goes through a **temporary productivity dip** (learning curve + verification tax + pipeline adaptation) before **exponential growth**, a metaphor for the *"tuition cost of transformation"* to be **budgeted explicitly**. Reference calculation: a 500 FTE organization / $176k fully loaded salary / 12.5% time saved per developer (≈ 1h/8h day) → **value $11.6M / investment $8.4M / ROI 39% / payback period 8 months (0.7 year)**. Modeled costs: licenses ($250/user/year), additional API costs ($80/user/year), training ($9,600/user/year), infra ($100k/year), J-Curve cost ($3.3M for a 15% drop over 3 months). Modeled value: **headcount reinvestment capacity** ($11M — freed capacity to reinvest, **NOT headcount reduction**), revenue from extra feature deployments ($990k, based on a 33% idea success rate, Larsen 2023), **negative downtime impact** (−$344k, "instability tax"). **Explicit reinvestment strategy**: ***"we strongly recommend organizations do not adopt a headcount-reduction strategy"*** — reinvest in innovation, retain talent, capitalize on institutional knowledge. Five pillars of value: Productivity / User Experience / Cost Efficiency / Developer Experience / Business Growth (from most direct to most indirect, *cumulated business value*). Five systemic keys to adoption: **Trust + Platform + Data + Users + Guardrails**. Two-phase roadmap: (1) **Build context layer (CapEx)** — quality IDP + healthy data ecosystems; (2) **Empower human in loop (OpEx)** — context engineering + trust in AI. Indicators: leading = experiment frequency + deployment frequency; stability gauge = change failure rate + rework. Three scenarios to model (Conservative 0.8 value × 1.5 cost / Realistic 1.0 / Optimistic 1.2 × 0.8). External data leveraged: 78% of executives report ROI on ≥ 1 gen AI use case (Google Cloud), 88% of early agentic AI adopters see positive ROI, **35-40% productivity greenfield vs ≤10% brownfield/legacy** (Stanford), inference cost ÷280 between Nov 2022 and Oct 2024 (Stanford AI Index 2025), **727% ROI over 3 years** for Google Cloud AI customers, average **8-month** payback in the AI market. Assumed weaknesses: *"all models are wrong"* — the model needs contextualizing, the calculator needs adjusting; risk of double-counting value (time saved → both avoided hire AND extra revenue); the user experience link is "loose" and therefore excluded from the calculator. **Deontological insight**: ***"We don't measure AI by the code it writes but by the bottlenecks it clears"*** — measured by bottlenecks cleared, not by code volume. **Major relevance** for CIOs/CTOs needing to build a defensible AI business case in front of a CFO/board; for France/Europe, to be read alongside Wescale (realistic X3-X4), Tatsyi/Raiffeisen Bank Ukraine (bank case study, −75 people but deliberate reinvestment), Frizzo (3-5× median), Curran/Intercom (3× R&D over 16 months), DORA Report 2025 (which this ROI report builds on).

The AI-native interview

Mon, 20 Apr 2026 00:00:00 GMT

In "The AI-native interview," Bret Taylor, co-founder and CEO of Sierra, describes the complete overhaul of the company's engineering hiring process to adapt it to the realities of AI-native development. The starting observation is clear: coding agents like Codex and Claude Code are upending software engineering. The engineer's role is no longer to "build the machine" but to "design and refine it." When a single engineer can build across the entire stack thanks to AI tools, competitive advantage comes from the combination of technical capability, product thinking, and business context — not from solving algorithmic puzzles. Sierra's old process was standard: two coding interviews, one algorithms interview, one system design interview, one culture-fit interview, then reference checks. The new process rests on three attributes: being representative of real day-to-day work, producing rich signal on the candidate's strengths and weaknesses, and offering a positive and authentic experience. The core of the overhaul is the new three-phase AI-native onsite. During the "Plan" phase, the candidate leads an ideation session to define a product to build, while evaluators ask questions to enrich the proposal. The idea is centered on the candidate's area of expertise to observe their product thinking in action. During the "Build" phase, the evaluator leaves the room and the candidate has two hours to bring their idea to life, using the AI tools and frameworks of their choice, with the freedom to pivot or adjust scope. Finally, during the "Review" phase, the candidate presents what they built: evaluators debate the product choices, examine the code to assess technical judgment, discuss the path to production, and explore how AI was used. Beyond the onsite, Sierra has replaced the coding phone screen with a system design interview, deemed more relevant for assessing the ability to ship code to production at scale. The company is also piloting a debugging interview in which the candidate receives a mid-sized codebase with a colleague's draft PR, and must review and improve it by iterating with coding agents. Evaluation criteria are agnostic to the product built, and interviews are conducted in evaluator pairs to improve calibration. Sierra explicitly hires for strengths rather than the absence of weaknesses. Candidate feedback has been enthusiastic: several have said it was "the most fun interview they've ever had."

IFTTD #351 - AWS Summit : Rester aux commandes des agents de code (avec Julien Lépine)

Wed, 08 Apr 2026 00:00:00 GMT

In episode #351 of *If This Then Dev*, **Julien Lépine** (CTO of AWS France) looks back at the **AWS Summit Paris** (April 1, 2026, ~10,000 attendees) and the transformation of the developer role. His angle: AWS is not merely a spectator but an **actor** in this change, driven by its own need — producing critical services at a dizzying scale (**DynamoDB: >5,000 billion requests/hour**; the October 20, 2025 outage that took Fortnite offline outside Europe). **Central proof point**: the redevelopment of **Amazon Bedrock**, AWS's agentic core. Estimated by Anthony Ligori (Distinguished Engineer) at **30 developers and 18 months**, it was completed by **6 people in 72 days** thanks to an **agentic platform** — **code entirely generated by AI**, largely reviewed, **without any vibe coding** given it is a critical production platform. As a result, AWS is **standardizing internally on Kiro** (IDE + CLI, running on Claude Sonnet/Opus, announced by Matt Garman at re:Invent) for ~30,000 developers, backed by a community (a Slack channel with 30,000 members, summarized every evening by AI) and shared **ADRs**. The central debate is about **value**: if one can "generate 100,000 lines of code per day," do best practices still matter? Lépine cites **Kent Beck** ("99% of my value became useless, but the remaining 1% multiplied by 1000"): value shifts toward **understanding context and architectural trade-offs**. Good practices (security, the maintainability principles of the **Well-Architected Framework**, rejecting **over-engineering**) remain, but the challenge becomes **keeping control without reviewing everything**: **formal TLA+ modeling** guaranteeing **invariants**, deterministic and mathematically proven **Raisonnement automatisé** to bound agents, with AI checking **code ↔ model divergence**. On **accountability**, the position is clear-cut: *"it is not the agent's responsibility, it is the responsibility of the person operating it"* — a culture of **blameless post-mortem** and **mechanism**. One incident (an agent with excessive permissions) led not to a shutdown but to new **guardrails**: any change impacting production, whether by a human **or** an agent, must be **reviewed** beforehand. **Regulated industries** (healthcare, defense, legal) adopt AI faster thanks to their existing data classification and their **auditability**. On the organizational side: **AI DLC** replaces sprints with multiple daily **Bolts**, AI absorbs **detailed specs**, PM/PO/Scrum Masters gain superpowers, and **communication barriers** fall. But **cognitive overload** looms: one client **deliberately slowed its pace** to protect its developers. Conclusion: **empathy, context, and understanding** become the key skills, and the line between tech and tech-adjacent roles is blurring.

Does LLMs / Vibe coding mean more or fewer developers?

Fri, 27 Mar 2026 00:00:00 GMT

Simon Wardley responds, in the form of a fictional Socratic dialogue, to the question of whether LLMs and vibe coding will lead to more or fewer developers. His answer is nuanced: probably about the same number, due to the Red Queen effect. As companies compete with one another, any productivity gain will be immediately reinvested to stay competitive. Concretely, a large company will go from 30 million to over a billion lines of code, simply to maintain its position. Wardley draws on the Jevons paradox as an explanatory framework: when a resource becomes more efficient, its consumption increases rather than decreases. Vibe coding makes code production cheaper, which will trigger an explosion in the volume of software produced, not a reduction in headcount. The central historical analogy is that of sysadmins. When virtualization made physical server racking obsolete, sysadmins did not disappear. They transformed into DevOps Engineers and SREs, acquiring new skills: chaos engineering, continuous deployment, distributed systems. Likewise, developers will not disappear but will evolve toward roles managing "packs of agents," making structural decisions, and maintaining a chain of understanding across increasingly complex systems. Wardley states that reading code is already unsustainable and that software engineering must transform from a craft into an engineering discipline to develop better methods of systemic understanding. The title "software engineer" will probably disappear — not because the role vanishes, but because too many executives have publicly declared these profiles no longer necessary and will not want to lose face. New titles will emerge: "Human-AI system integrator," "AI Wrangler," "Agentic Herder." The dialogue's punchline is biting. Wardley introduces "Alice," the developer laid off on a "thought leader"'s advice. Alice will soon be rehired, more expensive, under a new title. Managerial short-termism — laying off staff to boost stock options, then jumping ship before the consequences hit — is pinpointed as the real problem. Wardley recommends retraining over layoffs, while cynically predicting that companies will not take that path.

When Using AI Leads to "Brain Fry"

Thu, 05 Mar 2026 00:00:00 GMT

Six BCG researchers, including psychiatrist Gabriella Rosen Kellerman (*Tomorrowmind*), publish a study on March 5, 2026 in Harvard Business Review that gives the viral "AI fatigue" phenomenon its official name and measurement framework: ***AI brain fry***, defined as *"mental fatigue from excessive use or oversight of AI tools beyond one's cognitive capacity"*. Solid methodology: 1,488 full-time US employees, large companies, cross industries (January 2026). The article opens with two signals: the January 1 launch of **Gas Town** by Steve Yegge (orchestration of simultaneous Claude Code agent swarms) — *"Gas Town was moving too fast for me"* — and the viral X post by Francesco Bonacci (Cua AI) *"Vibe Coding Paralysis"*: *"I end each day exhausted—not from the work itself, but from the managing of the work."* The central finding empirically distinguishes **burnout** (emotional) from **brain fry** (acute cognitive). AI can **ease burnout** (-15% when it replaces repetitive tasks — *"toil"*) while **worsening brain fry** when it requires *intensive oversight*: +14% mental effort, +12% mental fatigue, +19% information overload among workers with a heavy supervision load. **14% of AI-using workers** report brain fry. Prevalence varies drastically by function: **Marketing 26%, HR 19%, Operations/Engineering 18%, Finance 17%, Legal 6%**. The productivity-tools curve plateaus at 3: 1 tool = 3.3 / 2 = 3.8 / **3 = 4.1 (peak)** / 4+ = 3.7. *Multitasking is notoriously unproductive, and yet we fall for its allure time and again.* Documented business costs: **+33% decision fatigue, +11% minor errors, +39% major errors, intent to leave 25% → 34% (+39% relative)**. Managerial practices: a manager who answers AI-related questions reduces fatigue by **-15%**. One who expects employees to figure it out on their own adds **+5%** — this is the ***"AI orphan tax"***. At the organizational level: "more work due to AI" = +12% fatigue; valuing work-life balance = **-28%** fatigue. Five recommendations for leaders: (1) holistically redesign jobs for shared human+AI responsibility, keeping neurobiology in mind; (2) set explicit expectations — *"70% of AI transformation efforts should be devoted to people and processes"*; (3) shift activity metrics toward impact; (4) develop workers' skills in **problem framing, analysis planning, strategic prioritization**; (5) treat human attention as a finite resource and evolve people analytics to monitor cognitive load. Pivotal 2026 academic piece, cited from April onward by Les Echos. It turns a Twitter buzz into a measured industry signal, and gives CHROs the quantified language to justify that the AI issue has now shifted from technology to the organization's cognitive governance.

It is amazing how many companies I talk to STILL have AI effectively blocked by IT & legal departments...

Thu, 05 Mar 2026 00:00:00 GMT

Ethan Mollick, associate professor at the Wharton School and author of Co-Intelligence, shares his astonishment at how many companies still have AI effectively blocked by their IT and legal departments, and for outdated reasons. He highlights the paradox: companies operating in highly regulated industries have already found ways to deploy enterprise AI solutions like ChatGPT, Claude and Gemini, including command-line tools like Claude Code, without encountering any apparent issue. Mollick describes what he calls one of the strangest gaps he observes: two companies in exactly the same industry can have radically different approaches. One has been using AI productively for 18 months, while the other has set up a committee that must approve every individual use case and still worries that current AI models automatically train on company data - a fear Mollick directly refutes by pointing out that this is not the case with enterprise versions. In a follow-up comment, Mollick deepens his analysis by identifying the root cause of this divergence. The decisive factor is generally the willingness of an executive - often the CEO, but not always - to take on the risk and responsibility associated with AI adoption. When the answer is no, the risk-reduction forces within the organization (IT and legal departments foremost) have every incentive to avoid anything that could even be suspected of causing a problem. His conclusion is clear: this is fundamentally a leadership issue. It is not a technical, regulatory, or data security problem - it is a matter of managerial courage and strategic willingness to embrace change. Companies that move forward are those whose leaders accept calculated risks, while those that stagnate are paralyzed by an institutional risk aversion that translates into bureaucratic approval processes and unfounded fears. This post generated massive engagement (over 1,793 reactions, 288 comments, 149 shares), reflecting how strongly this observation resonated with the professional community.

RAPPORT D'ANALYSE — ALMIA : La plateforme d'IA générative d'AG2R LA MONDIALE

Sun, 01 Mar 2026 00:00:00 GMT

Almia is the proprietary generative artificial intelligence platform developed in-house by AG2R LA MONDIALE, a major joint and mutual insurance group in France (15 million policyholders, 500,000 client companies). Launched among about a hundred "Champions IA" in April 2024, then opened to all employees in January 2025, the platform reached a remarkable adoption rate by the end of 2025: 7,000 users out of 8,300 employees, with roughly 1,300 AI assistants created by the users themselves. The platform is built around four complementary pillars. Almia Bot offers a secure chatbot that lets users query several LLMs via RAG, with an internal marketplace of assistants rated by users. Almia Apps industrializes the best-performing assistants into dedicated business applications (analysis of customer verbatims, marketing campaign generation, document processing). Almia Dev exposes generative AI APIs for developers and actuaries. More recently, Almia Agents deploys business AI agents integrated into the value chain. The infrastructure runs on S3NS, a Thales-Google Cloud joint venture hosted in France, qualified SecNumCloud 3.2 by ANSSI in December 2025. Data protection relies on vectorization: internal documents are broken down into unusable vectors before being sent to the LLMs, with the original texts remaining stored locally. The approach is multi-LLM and agnostic, allowing the best model to be chosen for each use case. The deployment strategy relies on a network of Champions IA drawn from the business units, each training their colleagues. Three mandatory training sessions (AI literacy, security, ethics) are required to access the platform. This peer-based approach, supported by the WEnvision consulting firm, enabled rapid and controlled adoption. A formalized AI governance body brings together the AI director, the DPO, the CISO, and CSR compliance to review all AI experiments and ensure compliance with the European AI Act. The principle is clear: AI assists employees without ever replacing human expertise. Almia is part of a six-year, €629 million IT transformation program. The project earned Pascal Martinez, IT and Digital director, the title of IT Strategist 2025. With the arrival of new CEO Fabrice Heyries at the end of 2025 and his "Esprit de conquête" (spirit of conquest) strategy, AI is now a strategic lever for differentiation for the group.

Fragments: February 13

Fri, 13 Feb 2026 00:00:00 GMT

Martin Fowler shares his reflections from the Thoughtworks retreat on the future of software development, an event bringing together industry experts to examine the impact of LLMs on development practices. First observation: senior developers are largely optimistic about LLMs. Their approach is to focus on architecture and treat AI agents as junior developers to be supervised. Notably, a third of initially skeptical seniors change their minds after hands-on exercises. Mid-level developers, by contrast, find themselves in a difficult position: their careers were built before the LLM era, yet they do not yet have the senior expertise needed to effectively orchestrate these tools. Margaret-Anne Storey introduces the concept of "cognitive debt," describing the situation where a team becomes unable to modify its code because it can no longer explain the underlying design decisions. Fowler draws a distinction between "cruft" — unintentional degradation through ignorance — and true technical debt, which involves a conscious choice and a calculated cost. Laura Tacho offers a striking observation: the Venn diagram between developer experience and agent experience is a perfect circle. Everything that makes work easier for human developers also makes it easier for agents. A telling paradox: leaders are willing to make accommodations for LLMs (documentation, code clarity, clean environments) that they stubbornly refused to make for their human teams. On the IDE front, the trend is moving toward a hybrid model combining non-deterministic tasks handled by LLMs and deterministic tasks like refactoring, opening up new orchestration possibilities. On team size, the consensus is that "two-pizza" teams will keep their size but increase their productivity. The question of pair programming with agents remains open and promising. Research published in the Harvard Business Review by Ranganathan and Ye offers an important counterpoint: AI adoption leads to work intensification and burnout. Initial productivity gains give way to quality degradation over the medium term. Camille Fournier sums up this tension with the phrase "everyone becomes a manager": supervised programming turns every developer into an agent manager, generating fatigue from constant context switching. This new paradigm demands supervisory skills more than direct execution skills.

2026 Agentic Coding Trends Report — How coding agents are reshaping software development

Sun, 01 Feb 2026 00:00:00 GMT

Anthropic publishes its report on agentic coding trends for 2026, identifying eight major developments that are redefining software development. The report is structured around three axes: foundational trends, emerging capabilities, and organizational impacts. The foundational trend (Trend 1) is the profound transformation of the software development lifecycle. Engineers are shifting from the role of implementers to that of AI agent orchestrators. Onboarding onto a new codebase collapses from weeks to hours, paving the way for dynamic "surge staffing." Augment Code illustrates this phenomenon: a project estimated at 4-8 months was completed in two weeks. On the capabilities side, four trends emerge. Single-agent systems are evolving into coordinated multi-agent teams (Trend 2), as at Fountain, where hierarchical orchestration reduced staffing for a logistics center from more than a week to less than 72 hours. Agents now operate on horizons of days rather than minutes (Trend 3) — Rakuten completed a full implementation in 7 autonomous hours on a 12.5-million-line codebase with 99.9% accuracy. Human oversight becomes intelligent (Trend 4): agents learn when to ask for help, and humans focus on strategic decisions. Finally, agentic coding extends to new surfaces (Trend 5), from legacy languages like COBOL to non-technical users. The organizational impacts are considerable. Productivity translates not only into speed but into increased output volume (Trend 6): 27% of AI-assisted work involves tasks that would never have been undertaken otherwise. TELUS saved more than 500,000 hours. Non-technical use cases are exploding (Trend 7): within Anthropic itself, the legal team reduced marketing review turnaround from 2-3 days to 24 hours, and Zapier reports 89% AI adoption with 800+ internal agents. Security presents a double-edged sword (Trend 8): the same capabilities benefit both defenders and attackers. The report emphasizes a central paradox: although developers use AI in 60% of their work, they fully delegate only 0-20% of tasks. AI is a constant collaborator requiring active oversight and human validation. The four priorities for 2026 are multi-agent coordination, scaling oversight, extending beyond engineering, and security-first architecture.

SoGPT "IA interne" vs Solutions du marché "Copilot" : le faux débat qui nous fait perdre la vraie bataille

Thu, 15 Jan 2026 00:00:00 GMT

Erwan Simon, CEO of GENIAL, reacts to Société Générale's decision to abandon its internal tool SoGPT in early 2026 to adopt Microsoft Copilot. He argues that this build vs buy debate constitutes a false framing that masks the real strategic question: who owns the business intelligence you build with AI? The article denounces the error of comparing ChatGPT or Claude to office software like Word or Excel. Unlike these finished products with fixed functionality, AI systems are both consumer interfaces and APIs enabling custom business applications. They are infrastructure, not commodities. Simon analyzes SoGPT's failure not as a validation of the buy model, but as an execution error: a generic chat interface disconnected from banking operations, without a continuous evolution strategy. The error was not building in-house, but creating a product with no operational anchoring. Copilot has its own limitations. Useful for generic tasks, it struggles with specialized operations such as ERP queries or compliance workflows. Many deployments plateau after the initial pilots due to difficulties measuring ROI, governance issues, and unintended data exposure via existing permission structures. The author cites AllianzGPT as a successful alternative model. Allianz built a platform, not just an interface, orchestrating multiple components: multiple language models (Azure OpenAI and Anthropic's Claude), connectors to internal systems, and full traceability of decisions for regulatory compliance. This architecture preserves organizational assets independently of any single vendor. Simon defines the concept of "AI capital" as the set of encodable business knowledge: documented processes, business rules, historical decisions, tacit expertise made explicit. It is this asset that creates a durable competitive advantage. The article concludes with a warning about European sovereignty. Europe lost cloud sovereignty in the 2010s by migrating to AWS, Azure, and Google. Repeating this mistake with AI means outsourcing the capacity to generate AI-driven value to American competitors. The fundamental question: will European companies build indigenous AI capabilities, or will they become permanent consumers of intelligence systems controlled by the United States?

Don't fall into the anti-AI hype

Sun, 11 Jan 2026 00:00:00 GMT

Salvatore Sanfilippo, better known by the pseudonym antirez and creator of Redis, publishes a position piece encouraging developers not to fall into the "anti-AI hype." Coming from a developer known for his commitment to minimal, well-written code, this testimony carries particular weight. The author acknowledges from the outset his social concerns regarding job losses and technological centralization linked to AI. However, he argues that "refusing what is happening now" will help neither programmers nor their careers. His central thesis: AI will change programming forever, regardless of ideological positions. To support his argument, antirez details four recent projects completed with Claude Code in hours rather than weeks. The first concerns modifying Linenoise, his command-line editing library, with the creation of a test framework emulating a terminal. The second involves debugging Redis failures related to TCP timing. The third example is particularly striking: a 700-line BERT en C implementation library generated in 5 minutes, with performance comparable to PyTorch. Finally, he mentions reproducing a design document for Redis Streams in 20 minutes. The author maintains a nuanced position. Despite these successes, he remains committed to the values of well-written, minimal code. He does not claim that AI replaces developer judgment, but that it considerably amplifies his ability to build. On the societal level, antirez recommends voting for governments that support those affected by automation, acknowledging that the transition will be painful for some. He nonetheless sees a beneficial potential of AI for science and innovation. His main message to developers: explore these tools "seriously" rather than rejecting them after superficial tests. He implicitly criticizes those who form their opinion on AI programming without thorough experimentation. The article concludes with a metaphor inviting readers to rediscover "the fire" of software building. AI is not there to replace this passion but to amplify it. Coming from such a respected figure in the open source community, this call for pragmatism rather than ideological rejection resonates particularly in current debates about the future of the developer profession.

2025: The State of Generative AI in the Enterprise

Tue, 09 Dec 2025 00:00:00 GMT

Menlo Ventures' third annual report on the state of generative AI in the enterprise documents unprecedented growth: enterprise AI spending rose from $1.7B in 2023 to $37B in 2025 ($11.5B in 2024, a 3.2x increase in one year), capturing 6% of the global SaaS market. This is the fastest-growing software category in history. The application layer dominates with $19B (51% of spending), ahead of infrastructure (33%) and model APIs (16%). The market is maturing: at least 10 products exceed $1 billion in ARR and 50 products exceed $100M. Adoption patterns are shifting markedly. Enterprises now buy 76% of their AI use cases rather than building them in-house (up from 53% in 2024). The production conversion rate reaches 47%, compared with 25% for traditional SaaS, a sign of clear perceived ROI. Product-led growth accounts for 27% of AI application spending (four times the rate of traditional software), and "Shadow AI" — employees paying for AI tools with personal cards — is estimated to account for nearly 40% of application spending. On the competitive front, AI-native startups capture 63% of the application market versus 37% for incumbents. This dominance is particularly pronounced by department: 71% in Product & Engineering (Cursor versus GitHub Copilot), 78% in Sales (Clay, Actively versus Salesforce), 91% in Finance & Operations (Rillet, Campfire versus Intuit). Incumbents, however, retain 56% of the infrastructure layer (Databricks, Snowflake, MongoDB). Departmental AI totals $7.3B (4.1x year over year), dominated by coding: $4B, or 55% of the total, with 50% of developers using AI coding tools daily and measured velocity gains of 15% or more. Next come IT ($700M), marketing ($660M), customer success ($630M), design ($490M), and HR ($370M). Vertical AI reaches $3.5B (2.9x), driven by healthcare ($1.5B, 43%, tripling since 2024), ahead of legal, financial services, education, and retail. The report settles the boom-versus-bubble debate in favor of a boom: strong real revenue growth, broad adoption (500+ decision-makers surveyed), measurable productivity gains, and sustainable business models. It also highlights the rise of model-agnostic tools (Cursor, OpenRouter) that accelerate adoption of frontier models, and international competition, notably from Chinese providers.

From Chalkboards to Chatbots: Evaluating the Impact of Generative AI on Learning Outcomes in Nigeria

Mon, 01 Dec 2025 00:00:00 GMT

The World Bank publishes the first rigorous study (RCT) evaluating the impact of generative AI on education in Sub-Saharan Africa. The intervention: a six-week after-school tutoring program using Microsoft Copilot (GPT-4) for English learning among first-year secondary school students in Benin City, Nigeria. **Transformative results**: The study demonstrates substantial improvements despite significant infrastructure constraints. The overall score increases by 0.31 standard deviation, English by 0.23 standard deviation (equivalent to 1.5 years of typical Nigerian schooling), AI knowledge by 0.31 standard deviation. Overall gains are equivalent to two years of schooling. A linear dose-response relationship shows that each additional day of attendance generates +0.031-0.033 standard deviation of improvement. **Notable differentiated effects**: Girls benefit from an additional 0.42 standard deviation effect, offsetting initial performance gaps. Students with higher baseline scores and those from more advantaged socio-economic backgrounds show larger gains, but disadvantaged students also achieve statistically significant improvements. **Exceptional cost-effectiveness**: At $48 per student for six weeks ($124 annualized), the intervention generates 3.2 equivalent years of schooling (EYOS) per $100 invested, surpassing most comparable educational interventions worldwide. The benefit-cost ratio reaches 161:1 to 260:1. Projected lifetime wage returns reach $7,767-$12,517 per participant. **Structured pedagogical approach**: The success relies on three days of teacher training, prompts designed according to learning science principles (retrieval practice, elaborative interrogation), awareness of AI hallucinations and biases, and active supervision of student engagement. The teacher acts as a "force multiplier" rather than being replaced. **Promising scalability**: The use of free software (no subscription), the absence of any need for proprietary question banks, and success with non-specialized staff suggest strong replication potential. The study addresses Bloom's "two-sigma problem": how to make the benefits of personalized tutoring accessible at the scale of entire populations in an economically viable way. **Critical context**: The study is set within the global learning crisis, where 70% of ten-year-olds in low- and middle-income countries cannot read a text appropriate to their level. These results position generative AI tutoring as a promising approach for resource-constrained contexts.

Moving away from Agile: What's Next? Reshaping Software Delivery with Agents

Sun, 23 Nov 2025 00:00:00 GMT

Martin Harrison and Natasha Maniar of McKinsey present a vision of software development transformation in the AI era, arguing that the traditional Agile model must evolve. Despite spectacular individual productivity gains from coding agents, many large enterprises plateau at an overall improvement of only 5 to 15%. This gap is explained by the emergence of new bottlenecks: manual code review that no longer keeps pace with generation speed, human collaboration ill-suited to the new cadence, and increased technical complexity. To unlock value, McKinsey identifies among "Top Performers" a transition toward **"AI Native" workflows and roles**: 1. **From "User Stories" to "Specs"**: Instead of iterating on vague textual descriptions, teams move to "Spec-driven development," where Product Managers (PMs) and developers iterate on precise technical specifications with agents, sometimes generating prototypes directly. 2. **Team reorganization**: The standard Agile team model (8-10 people) gives way to smaller, more autonomous "pods" (3-5 people, "One pizza teams"). Roles consolidate: less rigid specialization (separate QA, Frontend, Backend) and more "full-stack" profiles orchestrating agents. 3. **Continuous planning**: AI enables shorter planning cycles (from quarterly to continuous) and real-time roadmap adaptation. They present a case study of an international bank that reorganized its teams around specific workflows (bug fixing vs. greenfield) and used agents for task assignment and compliance verification, resulting in a 51% increase in code merges. The talk places heavy emphasis on **change management**. Technology alone is not enough; 70% of companies have not yet adapted their job descriptions. Success depends on a holistic approach including training ("upskilling"), redefining incentives (certifications, career paths), and rigorous measurement of impact (beyond mere adoption, looking at delivery speed, quality, and business outcomes). The future belongs to organizations capable of "rewiring" their operating model for symbiotic human-agent collaboration.