RFC processes fail at the review step: authors wait days for substantive feedback, then get a mix of typo nits and one architectural objection that should have arrived before they wrote section four. Reviewers, for their part, face a twelve-page document with no map of what's load-bearing.
The result is RFCs that are approved through fatigue rather than scrutiny — and decision records that don't capture why the decision was actually made.
A Fleet workflow gives every RFC a structured first review within minutes of submission — flagging unsupported claims, unconsidered alternatives, and missing failure analysis — then routes a flagged draft back to the author (bounded), and records the final human approval as the decision.
genflows:
- name: rfc-review
steps:
- {name: critique, prompt: "Review the RFC in the trigger ticket against the corpus: flag claims contradicted by existing design docs, alternatives dismissed without analysis, and missing failure modes.", corpus: ["docs/rfcs/**/*.md", "docs/architecture/*.md"], kind: review, out: critique.md}
- {name: approve, depends_on: [critique], kind: approval, out: decision.md}
In the builder, the review's flag outcome routes back to the author with a bounded retry count — a flagged RFC returns with specific objections, revised, and re-reviewed, at most N times before the gate escalates regardless. The human approver decides with the critique in hand, and the recorded decision becomes the RFC's audit trail.
Submission → structured critique against the existing design corpus → bounded revise-and-resubmit loop → human approval as the recorded decision. Reviewers spend their attention on judgment; the map of what's questionable is already drawn.
Teams with an RFC or design-doc culture where review latency and review depth are in visible tension — and where past decisions live in docs the next RFC should be checked against.
It replaces the first read, not the decision. The critique flags what deserves human attention — contradictions, gaps, unanalyzed alternatives — and the approval gate is always a named human. Most teams find human review gets sharper, because it starts from the flags.
Your accepted RFCs and architecture docs. That's the quiet compounding benefit: every accepted RFC strengthens the reviewer that checks the next one.
One binary. Five minutes. See every agent, coordinate every handoff, and keep a full audit trail of what your fleet did.