The Constraint Ring Around Agents in the Augmented SDLC
Episode « Phase 5 · Review » of SFEIR's series on the augmented SDLC, published the same day as Addy Osmani's LinkedIn post, which it translates into a phase specification.
By SFEIR// Source sfeir.com ↗/Reading 2 min/.md// Auto-verified translation
#ring of constraints#constraints around agents#Review phase#phase 5#augmented SDLC#SFEIR 11-phase cycle#human gates#Define Plan Ship
Summary
Fifth episode of SFEIR's series on the augmented SDLC, devoted to the Review phase, and published the same day as Addy Osmani's LinkedIn post, which it converts into a phase specification.
The starting observation: quality used to be read in the code; agents now produce more of it than anyone can review. It has therefore changed address — it lives in the ring of constraints surrounding the agent, that is, in the harness. Seven dimensions make up this ring (correctness, security, performance, accessibility, maintainability, economic efficiency, understandability), linked by the back-pressure rule: a loop is granted only the autonomy one knows how to verify cheaply and reliably. The corollary overturns the prevailing intuition: the bottleneck was never generation, it is verification — "generation is a wide mouth, verification a narrow neck; speeding up the mouth thickens the pile at the neck."
Phase 5 · Review
— SFEIR , sfeir.com
Hence the central architecture decision: in the eleven-phase cycle, Review is not a human gate, and this is deliberate. The three inviolable gates are Define, Plan, and Ship. Putting the gate on Review would place human attention — a finite resource — as the checkpoint of a generation that, itself, scales: the neck would never widen. Review instruments, Ship decides; Review produces an opposable body of evidence, and the decision is made on the evidence, not on the full diff. SFEIR keeps from Monperrus that human inspection of every diff does not withstand agentic speed, but rejects his conclusion: acceptance cannot be delegated.
The operational translation is a dimension-by-dimension table, separating what can be mechanized from irreducibly human judgment. The dimension systematically forgotten is understandability, "because it doesn't break CI" — hence the cheapest remedy in the grid: having the agent log what it tried and discarded, since "intent is not lost, it is thrown away."
The named failure mode is circular validation: the agent that writes the code writes the tests that validate it, CI is green, "you have built a mirror, not a ring." Five countermeasures are borrowed from Anthropic (independent gates, deterministic + agentic, shadow mode, risk-based tiering, SIEM logging), and Compare the Market warns that a reviewer built on vector RAG degrades the review (~70% for an AST graph versus ~58%).
The firm's own extension is the ratchet, attached to Compound-1: every escape becomes a constraint. The ring thickens with each cycle — "the only asset in the chain that appreciates while models depreciate" (− 30% fewer correction iterations after ten cycles, an internal measure). Only one question remains: what does my system refuse to let through?
Key takeaways
Nature of the note. this is a framing article, not a primary source of results. The text's own value comes down to three points, everything else is sourced synthesis: (1) Review outside the human gates and the split "Review instruments / Ship decides"; (2) attaching the ratchet to Compound-1 as the phase in charge of it; (3) the firm's own figure of − 30%. Do not re-cite Anthropic's or Compare the Market's figures via this note: go to the source notes.
⚠️ Two Osmani sources cited are NOT in the corpus and are the best candidates to add after this note: Set the constraints around your agents (LinkedIn, July 30, 2026 — this is the original diagram) and especially Software Factories, Light and Dark (addyosmani.com, July 2026), which alone carries three of the structuring concepts taken up here — the back-pressure principle, operationalized comprehension debt, and the workable length of loops (3 to 10 steps, losing the thread beyond about twenty). The corpus contains comprehension debt via [[osmani-cognitive-surrender-comprehension-debt-2026-05-05]] and the ratchet via [[osmani-agent-harness-engineering-2026-04-19]], but not the essay that links them to the lit/unlit factory metaphor.
⚠️ Citation variance to be aware of: SFEIR references Monperrus under the title "The End of Code Review: How AI Agents Supersede Human Code Review." The title carried by the corpus note (arXiv 2606.13175) is "The End of Code Review: Coding Agents Supersede Human Inspection." Same paper, same date (June 11, 2026); it is SFEIR's wording that drifts. Use the arXiv title for formal citation.
The dimension-by-dimension grid is reusable asset #1. — a two-column table directly transposable into a Review-phase specification, dimension → mechanizable constraint → residual human judgment:
Correctness. unit, property, mutation testing, green/red oracle → functional acceptance
Understandability. the agent logs what it tried and what it discarded, a decision log attached to the PR → reconstructing intent
Economic efficiency. token/compute budget per task, cost per change → TCO and the CapEx/OpEx trade-off
Back-pressure. (the rule linking all seven): autonomy ≤ low-cost verifiability → where to place the switch
The strongest architecture argument, reusable verbatim in a committee. if Review carried the human gate, "the system's checkpoint would be human attention, a finite resource that doesn't scale, facing a generation capacity that, itself, scales" — so the neck would never widen, and "you would have built a pipeline whose maximum throughput is the number of diffs a senior can read before the end of the day." This is the argument that justifies moving the gate to Ship rather than removing the human.
Operative distinction to remember. Review has a deliverable (an opposable body of evidence), Ship has a decision — "and that decision is made on the evidence, not on the full diff."
The forgotten dimension, and why.understandability is systematically omitted "because it doesn't break CI." The remedy is the cheapest in the grid and the least applied — asking the agent to write down what it tried and discarded, since "reviewing an agentic PR is the first time a human reconstructs the why." Phrase worth keeping: "intent is not lost, it is thrown away."
Circular validation, a failure mode to name in front of a team. the agent writes the code, the same agent writes the tests, CI is green → "a mirror, not a ring." Term borrowed from Augment Code ([[hingel-augment-how-ai-changes-sdlc-six-stages-2026-06-08]]); the measurement of the phenomenon comes from DORA 2025 (AI adoption correlated positively with throughput and negatively with stability when the foundations don't keep up). Diagnostic question #2 of the article, the most profitable of the five: "are your tests written by the agent that writes the code?"
Qualifying criteria for an autonomous loop. (reusable five-item checklist): the control must be cheap, high-frequency, hard to bypass, immediate, and non-drifting. Qualifying examples: green/red oracle, type gate, property tests, an agentic reviewer equipped with a real rubric. Corollary not to lose: short loops verify better than long ones (an agent holds up for 3 to 10 steps, loses the thread beyond about twenty) — "a loop that sprawls hides its errors in the corners."
Where to keep the light on. subtle bugs invisible to tests, wide blast radii, decisions structuring a year of work — namely authentication, billing, public API contracts. And the governance warning: "the real risk is setting every switch the same way" (all off → dismantled four months later; all on → no more shipping).
The ratchet, SFEIR's own extension. "every escape becomes a constraint" — a defect that has crossed the ring is not just fixed in the code, it closes back up inside the ring (test, lint rule, review rubric, harness guardrail), at Compound-1, with memory reloaded at the Plan phase of the next cycle. This is Osmani's ratchet principle (every error becomes a rule, cf. [[osmani-agent-harness-engineering-2026-04-19]]) wired to a phase of the cycle explicitly in charge of it — this, and nowhere else, is where SFEIR's framework adds something to the borrowed diagram. Two phrases: "a checklist is written once and goes stale, the ring thickens with every cycle" and "a bug seen twice is not a bug, it's a hole in the system."
⚠️ Figure to handle with caution: − 30% fewer correction iterations after ten cycles is labeled Measured · SFEIR, 2026 first-party material — no protocol, scope, sample size, or definition of "correction iteration" is given. It is the article's only own figure and it supports its most commercially useful thesis (the ring as an appreciating asset). Do not reuse it with clients without qualifying it.
Economic aphorism worth keeping. the ring is "the only asset in the chain that appreciates while models depreciate" — an investment argument (the harness survives model changes) to be linked with harness engineering doctrine.
Economic efficiency in a quality grid. the point is theoretically interesting and argued in a single sentence — "a per-task token budget is a constraint on the same footing as a performance budget, and it produces the same virtue, bounding autonomy by the cost of verifying it." This is the FinOps ↔ quality junction, complementary to the "risk-based tiering = FinOps policy" angle in the July 26 note.
Metadata. published on July 30, 2026, unsigned (firm voice), site tags sdlc, ia-agentique, software-factory, harness-engineering, code-review. Eight numbered sources with status labels. Published the same day as the commented Osmani post — a remarkable turnaround, worth considering a tech-watch fact in itself (the firm follows Osmani's LinkedIn thread in real time and converts it into a series episode). The article references an SFEIR text absent from the corpus: Code review in the age of AI: from creator to verifier (April 1, 2026) — candidate for addition, already flagged in the July 26 note.
Tech-watch dossier connections.
In-house framework, read first. [[sfeir-sdlc-ia-cycle-11-phases-2026-06-16]] (the 11 phases and the 3 gates — essential for situating "phase 5, between Verify and Compound-1") and [[sfeir-sdlc-pdlc-articulation-2026-07-22]].
Twin note, same week. [[sfeir-anthropic-sdlc-ai-native-securise-2026-07-26]] — the five anti-circularity mechanisms taken up here are its subject; this note carries the phase specification, that one the condition of possibility of the cycle.
Source of the Anthropic figures. [[clinton-anthropic-secure-ai-native-sdlc-2026-07-21]] (16% → 54%, one-third of incidents, shadow mode, risk-based tiering, SIEM).
Osmani, conceptual foundation. [[osmani-agent-harness-engineering-2026-04-19]] (harness, ratchet principle) and [[osmani-cognitive-surrender-comprehension-debt-2026-05-05]] (comprehension debt, ~100-line PRs, junior-standard review) — the ring's "understandability" dimension is exactly the structural guardrail that Cognitive Surrender called for.
Antagonistic position taken deliberately. [[monperrus-end-of-code-review-agents-supersede-2026-06-11]] — SFEIR takes the diagnosis, rejects the conclusion.
Neighboring framework, to compare on gate count. [[williams-adlc-2-two-human-gates-2026-06-12]] (two human gates) versus SFEIR's three; [[williams-adlc-4-prosecution-not-code-review-2026-06-12]] (review as a prosecution brief, the same shift toward evidence); [[williams-adlc-3-tests-are-the-spec-2026-06-12]] (the test as a language the builder cannot argue with — Osmani's phrase about "constraints the model cannot argue with" is its counterpart); [[williams-adlc-6-lifecycle-gets-cheaper-2026-06-12]] (the cycle that improves with every run = the ratchet).
Circular validation and the mirror effect. [[hingel-augment-how-ai-changes-sdlc-six-stages-2026-06-08]], [[dora-report-2025-ai-software-dev-2025-09-23]], [[dora-google-cloud-roi-ai-assisted-software-development-j-curve-2026-04-21]].
Context for an AI reviewer. [[comparethemarket-context-retrieval-ai-code-review-gkg-rag-2026-03-06]] — the source of the ~70% (AST graph) vs ~58% (RAG) figures over 79 MRs; to read before building an agentic review on vector RAG.
Factory metaphor. [[mccarthy-strongdm-software-factory-agentic-moment-2026-02-06]] and [[wescale-usine-logicielle-augmentee-juge-strategique-2026-05-03]] — "the lights-out factory" comes from Software Factories, Light and Dark.
"the structural knowledge graph built via AST analysis places a relevant inline comment in about 70 % of cases versus 58 % for vector RAG, across 79 merge requests"
software quality is no longer read in the code but in the ring of constraints surrounding the agent
— SFEIR
the bottleneck was never generation but verification: accelerating generation only thickens the pile at the verification neck
— SFEIR
"Review instruments. Ship decides." — Review delivers a contestable body of evidence, Ship decides on the evidence rather than the full diff
— SFEIR
"a loop is granted only as much autonomy as can be verified reliably and at low cost — not an inch more"
— Addy Osmani
placing the human gate on Review would make human attention the control point for a generation that scales, capping throughput at the number of diffs a senior can read in a day
— SFEIR
The knowledge graph extracted from this fiche — 9 entities, 30 relations.
In this graph :anneau de contraintes · back-pressure · phase Review (SDLC) · validation circulaire · cliquet de l'anneau · comprehension debt · Addy Osmani · Software Factories, Light and Dark · cycle SFEIR à 11 phases