Skip to content

SYSTEM Cited by 1 source

Cloudflare incident-report reviewer

Cloudflare incident-report reviewer is a Codex-backed agent that reviews incident reports (postmortems) for completeness and corrective-action quality. It applies a dedicated Codex RFC to assess whether the report explains what happened, identifies contributing factors, records the resolution, and proposes meaningful follow-up actions. (Source: sources/2026-08-04-cloudflare-how-cloudflare-enforces-engineering-standards-using-ai)

Architecture and role

The system shares the same Developer Platform building blocks as Cloudflare's spec reviewer: a Worker-hosted application, D1-persisted state/results, and model requests routed through AI Gateway. The shared substrate makes the Codex a reusable standards source rather than a code-review-only prompt.

Unlike the spec reviewer, this agent evaluates a retrospective artifact. Its value is therefore not real-time mitigation; it detects missing timelines, incomplete detection signals, omitted action items, and insufficient causal or resolution narratives while the incident learning is still actionable.

Operational evidence

Since May 2026, the reviewer has assessed more than 200 incident reports. Cloudflare reports that 93% covered low-impact, internal-only, or preemptively declared incidents. For high-severity incidents, the review is mandatory in the central process and a report is not complete until its findings are addressed. (Source: sources/2026-08-04-cloudflare-how-cloudflare-enforces-engineering-standards-using-ai)

Design significance

This system closes the loop between operational learning and future enforcement:

Incident → complete report + follow-up actions → Codex/RFC refinement
         → standards retrieved by design/code agents → earlier prevention

That loop operationalizes institutional memory beyond a postmortem archive: the same standards corpus informs both the quality of the report and the later reviews intended to prevent recurrence.

What is not disclosed

The post does not name the dedicated RFC, the report source system, report schema, model(s), latency, review-override mechanism, severity threshold for mandatory review, or quantitative false-positive/false-negative rates.

Seen in

Last updated · 622 distilled / 1,953 read