# davis-gill-hendrickson-itrevolution-agentic-ai-code-reviews-2026-09-30

## Veille

A collective article by seven authors (**Zach Davis**, **Michelle Gill**, **Elisabeth Hendrickson**, **Angie Jones**, **Sha Ma**, **Randy Shoup**, **James Wickett**), published in the *Enterprise Technology Leadership Journal* by **IT Revolution** (fall 2026, vol. 3, no. 2), under a CC BY-NC-SA license. About **7,500 words**, titled *« Toward a Pattern Language for Code Reviews in the Age of Agentic AI »*. **(A)** Framing: code review is defined as an examination by a non-author entity leading to an action; empirical studies (Microsoft 2013, Mäntylä 2009, Beller 2014) show that its value lies mainly in maintainability and knowledge sharing, less in functional defect detection. **(B)** Three conditions for taking the human out of the critical path: maximal deterministic validation, **risk-graded** review, a **durable trace** of every agent decision; *« no human in the loop »* is a configuration choice, not a default. **(C)** Nine composable patterns, presented as a toolbox rather than a maturity ladder: small changes, Shift Left, design for verification, deterministic engineering practices, context engineering, continuous quality loop, adversarial LLM review, multi-agent reviews, specification-driven development. **(D)** Three case studies: **Block**, **Cloudflare** (131,246 reviews across 48,095 PRs in one month, 1.2 findings per review), **DryRun Security**. Extends the corpus's notes on context engineering and agent-assisted code review.

## Titre Article

Agentic AI and Code Reviews: Toward a Pattern Language for Code Reviews in the Age of Agentic AI

## Date

2026-09-30

## URL

https://reader.itrevolution.com/download/agentic-ai-and-code-reviews?via=itrevolution

## Keywords

agentic code review, code review, pattern language, pattern, adversarial review, multi-agent review, continuous quality loop, shift left, deterministic verification, risk-graded review, context engineering, AGENTS.md, specification-driven development, small changes, stacked PRs, audit trail, human in the loop, Zero Trust, structural independence, Block, Cloudflare, OpenCode, DryRun Security, IT Revolution, dark software factory

## Authors

Zach Davis (LaunchDarkly), Michelle Gill (Engineered by AI), Elisabeth Hendrickson (Curious Duck), Angie Jones (Agentic AI Foundation), Sha Ma (Topogy), Randy Shoup (CircleCI), James Wickett (DryRun Security) ; article collectif, IT Revolution.

## Ton

**Profile**: a practitioner-research paper aimed at technical leaders, published in an editorial journal of the DevOps ecosystem. The authors come from tool vendors (LaunchDarkly, CircleCI, DryRun Security, Topogy) and consulting; the journal notes that AI tools may have helped with the writing.

**Style**: structured into patterns with a constant template (problem, applicability, mechanism, human role, links to other patterns), referenced to Christopher Alexander. General, normative sentences, few figures outside the Cloudflare case.

**Epistemic position**: a proposed framework, not evaluated through measurement. The case studies are written by or with the companies involved; only Cloudflare provides volumes. The article itself names the limits of AI (architectural judgment, cross-system impact, subtle concurrency issues, very large changes). Two authors run vendors of review or verification products, as disclosed on the authors page.

## Pense-betes

- **Code review definition**: examination of code by a non-author entity, resulting in a judgment followed by an action (merge, rejection, feedback). Distinguishes review from casual reading; the entity can be an agent.
- **Empirical value**: review is likely overrated as a defect detector and underrated as **risk triage**; most findings concern evolvability (maintainability, style, sharing).
- **Three conditions** for taking the human out of the critical path: (1) deterministic first (tests, static analysis, policy as code), the agent only handles the subjective residue; (2) **graded** review from "no human" to explicit authorization; (3) **inspectable record** of what was evaluated, under which policies, with what confidence, by whom.
- **Placement in the cycle**: during authoring (fast feedback), pre-merge (policy, architecture, regressions), post-merge (system drift), pre-production (release governance).
- ⭐ **Continuous quality loop**: an author agent and a reviewer agent in a six-step cycle (production, examination, structured feedback, correction, re-evaluation, stop on criterion: clean pass, max iteration count, or acceptable residual risk). The reviewer doesn't duplicate what linters do; it must **verify that the fix worked** and not look for new problems once the old ones are resolved.
- **Adversarial LLM review**: structural independence between generation and evaluation. Three forms: cross-model, cross-session (same model, isolated context, critical prompt), structured adversarial prompting. Analogies with GANs and Zero Trust (explicit verification, least privilege). Reserve for sensitive paths: cost and noise.
- **Multi-agent reviews**: specialists (security, performance, compliance, standards, correctness, exploitability), a triage agent, a **coordinator** (deduplication, severity, blocking/advisory policy), context routing. Cost grows with the number of specialists; the coordinator's quality determines the level of trust.
- **Context engineering**: the agent needs the architecture, API contracts, incidents, past review comments, accepted trade-offs. Without it: noisy, superficial findings. The human role becomes that of **curator**.
- **Specification first**: shift the control point from implementation to the plan; the human reviews the intent, the adversarial agent verifies fidelity to the specification.
- **Cloudflare**: CI-based orchestration around OpenCode, specialized reviewers and a coordinator, graded review, filtering of lockfiles and generated artifacts, a policy biased toward approval, a human "break glass" option, findings carried forward between commits. **131,246 reviews, 48,095 PRs, 5,169 repositories, 159,103 findings** (March-April 2026).
- **Block**: 3,500 engineers, 90% using AI daily by mid-2025; agent PRs stay in draft for lack of ownership; noisy in-house reviewers, then more useful commercial tools; stacked PRs, local adversarial review before push.
- **DryRun**: a multi-tenant isolation flaw in beta (2023) turned into a code policy; thirty-second feedback; exploitability test.
- ⚠️ **Limitations**: framework unmeasured outside Cloudflare; vendor case studies describing their own practice.

## RésuméDe400mots

The article proposes a pattern language for code review in the age of agents. It starts from an observation: agents accelerate the volume of changes, senior engineers receive dozens of review requests per day, and exhaustive human review is no longer sustainable. Code review is defined as the examination of code by an entity that is not its author, resulting in a judgment followed by an action. The empirical studies cited show that its value lies as much in maintainability, the discovery of alternatives, and knowledge sharing as in bug detection; it would mainly be a risk-triage mechanism.

To take the human out of the critical path, three conditions are set. A maximum amount of validation must be shifted to deterministic checks, so that agents only handle the residual subjective space. Review must be graded by risk, from no human at all to explicit authorization. And every agent decision must leave a durable trace of what was evaluated, under which policies, with what confidence. Working without a human in the loop is a configuration choice justified by risk and evidence.

Nine patterns make up the toolbox: small incremental changes, Shift Left, design for verification, deterministic engineering practices, context engineering, continuous quality loop, adversarial LLM review, multi-agent reviews, specification-driven development. The continuous loop has an author agent and a reviewer agent exchange until a convergence criterion is met. Adversarial review provides structural independence, through a different model or an isolated session. Multi-agent reviews split the work among specialists, a triage agent, and a coordinator. Context engineering conditions the rest: without it, findings are noisy.

Three cases illustrate this. Block, faced with agent PRs left unreviewed, moved from noisy in-house reviewers to commercial tools, stacked PRs, local adversarial review, and automatic correction in CI. Cloudflare built a CI-based orchestration around OpenCode: specialized reviewers, a coordinator, graded review, filtering of noisy files, a human bypass, and carrying findings forward from one commit to the next. Over one month, 131,246 reviews cover 48,095 PRs across 5,169 repositories, with an average of 1.2 findings per review. DryRun Security turned a tenant-isolation flaw into a policy verified on every change.

In conclusion, the authors present the transition as a reorganization of responsibilities: precise specifications, independent verification, durable audit trails. The guiding question: if something went wrong, how would anyone know?

## GrapheDeConnaissance

- Zach Davis —travaille_chez→ LaunchDarkly (ORGANISATION, 0.9)
- Randy Shoup —travaille_chez→ CircleCI (ORGANISATION, 0.95)
- James Wickett —dirige→ DryRun Security (ORGANISATION, 0.95)
- IT Revolution —publie→ article Agentic AI and Code Reviews (DOCUMENT, 0.95)
- article Agentic AI and Code Reviews —recommande→ revue graduée par le risque, avec trace durable de chaque décision d'agent (AFFIRMATION, 0.93)
- article Agentic AI and Code Reviews —affirme_que→ « No human in the loop » est un choix de configuration justifié par le risque et la preuve, non une valeur par défaut (AFFIRMATION, 0.92)
- revue adverse LLM —améliore→ l'indépendance entre génération et évaluation, par un modèle ou une session distincts (AFFIRMATION, 0.9)
- revue adverse LLM —s_inspire_de→ Zero Trust (CONCEPT, 0.88)
- boucle de qualité continue —utilise→ revue adverse LLM (METHODOLOGIE, 0.8)
- revues multi-agents —utilise→ agent coordinateur (CONCEPT, 0.92)
- ingénierie de contexte —permet→ des constats de revue moins bruyants et plus pertinents (AFFIRMATION, 0.9)
- développement piloté par la spécification —s_applique_à→ la revue de l'intention avant l'écriture du code (AFFIRMATION, 0.9)
- Cloudflare —a_créé→ système de revue de code en CI (TECHNOLOGIE, 0.93)
- système de revue de code en CI —utilise→ OpenCode (TECHNOLOGIE, 0.95)
- système de revue de code en CI —mesure→ 131 246 revues sur 48 095 PR dans 5 169 dépôts, 159 103 constats, soit 1,2 par revue (mars-avril 2026) (MESURE, 0.95)
- Block —utilise→ revue adverse locale avant push (METHODOLOGIE, 0.88)
- Block —observé_dans→ PR d'agents laissées en brouillon faute de propriété (AFFIRMATION, 0.85)
- DryRun Security —utilise→ test d'exploitabilité (METHODOLOGIE, 0.9)
- article Agentic AI and Code Reviews —référence→ étude Microsoft sur la revue de code moderne (DOCUMENT, 0.9)
- article Agentic AI and Code Reviews —affirme_que→ « Generating code is cheap, but organizational attention is not. » (CITATION, 0.85)

---
Canonical: https://www.thekb.eu/en/fiches/davis-gill-hendrickson-itrevolution-agentic-ai-code-reviews-2026-09-30/