Darragh Curran, R&D leader at Intercom, published on April 16, 2026 on Fin Ideas an unprecedented report card in the 2026 corpus. Nine months earlier, he had publicly committed to doubling R&D productivity in 12 months through AI. Result: 3x over 16 months, with no signs of plateauing. The article documents the trajectory with rare quantitative transparency.
Intercom's scale gives the case its significance: 500 people in R&D, 8.5 million lines of code, 313 production deployments per day, 30,000 B2B customers. The pivot metric is the merged PR, treated as a throughput that exposes bottlenecks. R&D is explicitly viewed as a "factory for producing high-quality increments". The strategic decree: "All technical work is becoming agent-first. This is the top priority for R&D."
The numbers: defect backlog -54%, product changes >2x, time idea→shipping -39%, breaking-changes downtime -35%, cost per PR -50%. 93.6% of PRs are agent-driven, 19.2% AI-approved (target >50%). 497 autonomous PRs in the first 4 weeks. Top 5% of performers: 6x the median PR throughput — "token spending correlates with individual gains". Claude Code in auto-approved merge mode at 14.6 min vs 75.8 min for the organization's median.
The key organizational innovation is the Skills-Based Plugin Architecture: a private marketplace distributing specialized Claude Code configurations, with auto-update. 153 contributors created 267 skills in 3 months, 31% of R&D actively contributes. The plugin ecosystem was the subject of a viral thread by Brian Scanlan. Outside R&D: 1,100 Claude Code users out of 1,305 employees — finance, recruiting and sales build their own analytics tools (Cormac platform, Streamlit-on-Snowflake).
Curran proposes a Productivity Tiering across 5 dimensions (AI usage intensity, output, depth, $/PR, prompt quality) to identify progression from minimal to elite. Anti-paralysis methodology: "don't search for perfect measures, embrace imperfect ones + monitor holistic outcomes". A following series announces the 2x Principles, and a public webinar is scheduled for May 19, 2026.
The Intercom case rounds out Stripe Minions, StrongDM Software Factory and Anthropic's 16-Claude compiler as a quantitative reference piece on agent-first at scale. It empirically validates Karpathy's theses (peaks well beyond 10x for the good ones), confirms the AI-approved → 50%+ trajectory (Sierra), and demonstrates that a B2B scale-up can triple its output with the same team — without triggering the permanent underclass dystopia (Sun NYT). Rare public receipts.
Key takeaways
Date / source. April 16, 2026, Fin Ideas (Substack, Intercom's media platform). Author: Darragh Curran, Intercom R&D leader.
Prior commitment (July 2025). Curran had publicly committed Intercom to doubling R&D productivity in 12 months through AI. It was a rare and verifiable C-level bet.
Result 9 months later.3x over 16 months, "no signs of plateauing" — target exceeded.
Intercom's scale.
500 people. in R&D
~8.5 million lines. of application code (multi-language)
2+ million QPS. at peak
313 production deployments / day.
30,000 business customers.
1,305 employees. total
1,100 Claude Code users. at peak (the entire company — finance, recruiting, sales build their own analytics tools)
Pivot metric.merged PRs as a throughput metric that exposes systemic bottlenecks. Pressure applied on this unit.
Factory model doctrine. R&D viewed as "factory for producing high-quality increments" — an assumed lean / DORA heritage.
Central principle."All technical work is becoming agent-first. This is the top priority for R&D." Top-down strategic decree.
Results table. | Metric | Result | |----------|----------| | Defect backlog | -54% | | Product changes | >2x | | Time idea → shipping | -39% | | Code quality | 5 weeks of positive progression | | Breaking-changes downtime | -35% | | Cost per PR | -50% | | Agent-driven PRs | 93.6% | | AI-approved PRs | 19.2%(target >50%) | | Autonomous PRs (first 4 weeks) | 497 | | Active plugin contributors | 31% of R&D | | Claude Code median merge time (auto-approved) | 14.6 min vs 75.8 min (org median) |
Uneven productivity."the top 5% of performers generate 6x the median PR throughput". Token spending correlates directly with individual productivity gains. This is the Karpathy pattern "10x is not the speed up — people who are very good at this peak a lot more than 10x".
Productivity Tiering. (Curran framework): 5 evaluation dimensions to identify progression from "minimal to elite agentic tool usage": 1. AI usage intensity 2. Overall output 3. Usage depth 4. Cost efficiency ($/PR) 5. Prompt quality
Private marketplace. distributing specialized Claude Code configurations across the entire organization
Auto-updating plugins. for rapid capability scaling
153 contributors creating 267. specialized skills in 3 months
31% of R&D actively contributes. to the marketplace
The plugin ecosystem is the standout angle — Brian Scanlan (Intercom member) published a viral thread on the topic
Flagship internal use case — Cormac (data analytics platform). February prototype → widespread adoption, Streamlit-on-Snowflake deployments extended to several departments (finance, recruiting, sales).
2x Principles. (announced but reserved for the next post): describe Intercom's "modern work methodology". Curran announces a multi-part series on the "messy journey", lessons learned, future roadmap.
Public webinar announced.May 19, 2026, 9am PT / 5pm GMT, targeted at organizational leaders seeking AI transformation strategies.
Anti-analytical-paralysis method."Don't paralyze decision-making searching for perfect measures." Embrace imperfect measures + holistic outcome monitoring. Wise for 2026, when many organizations remain stuck on the measurement problem.
Auto-approval methodology. a dedicated post is planned on AI auto-approval methodology and risk mitigation. Intercom is currently at 19.2% AI-approved, targeting >50% — a massive leap requiring governance trust and a proven risk model.
External mentions.
Ramp. comparable example of an organization pursuing similar outcomes (external validation of the pattern).
Brian Scanlan. (Intercom): viral thread on the plugin ecosystem.
Cormac. (Intercom): team lead of the data analytics platform.
Claire Vo. draft reviewer, thanked.
Anthropic. Claude / Claude Code provider (the core technical stack).
Watch-list dossier articulation.
Quantitative reference piece. on agent-first at scale-up scale. Comparable to:
16 parallel Claude agents C compiler. (Carlini/Anthropic 2026-02-05).
Confirmation of Karpathy's thesis."the speed up is not 10x, it peaks much higher for those who are good" (2026-04-29): Intercom measures exactly this pattern (top 5% at 6x median).
"Agent-first" decree. corroborates Levie Building for Trillions of Agents (2026-03-07), Greyling CLI vs IDE (2026-03-09), Wescale Augmented Software Factory (2026-05-03).
Skills-Based Plugin Architecture. = operational implementation of Vincent Superpowers (2026-04-02), Anthropic Skills (2025-10-16), Karpathy Skills for Claude Code (2026-01-27).
AI-approved PRs 19.2% → 50% target. = concrete trajectory of Sierra's AI-native interview thesis (Taylor 2026-04-20) on the production side: if the agent can auto-approve 50% of PRs, that's a fundamentally transformed quality gate.
Imperfect metrics acceptable. aligns with Stanford quantify AI ROI (Denisov-Blanch, 2025-11-23), Reock DX Leadership AI Engineering Metrics (2025-11-23) — moving past the measurement debate into action.
Cross-functional Claude Code adoption (1100/1305). validates Mollick's Real AI Agents (2025-09-29), Levie's Building for Trillions of Agents (2026-03-07) theses on expansion beyond tech.
Receipts of a public commitment kept. to be put in perspective with the 50% of jobs gone by 2030 from Amodei (Sun NYT 2026-04-30) — Intercom did not lose 50% of R&D, but tripled output with the same team, which validates the thesis without triggering the dystopia. Classic productivist trade-off.
Limitations to flag. no independent external review of the figures, no comparison with a rigorous pre-AI baseline other than the internal trajectory, defect backlog -54% may also reflect parallel non-AI cleanup. But the amplitude of the metrics and the consistency of the ecosystem (plugins, productivity tiering, auto-approval, cross-functional usage) make the trajectory credible.
To leverage for. CFO/CEO business case on R&D transformation; CTO argument for pushing agent-first; benchmarks to present at French COMEX; defining agent-first KPIs (cost/PR, AI-approval rate, agent-driven %, plugin contribution %); designing an internal private skills marketplace.