The agent that improves the agents.
It studies every finished file, finds what slowed it down, and proposes the fix — reviewed by a human before anything ships.
What it is
A credentialing operation generates a rich record: every requirement, every tool call, every wait, every human correction. The Evolution Agent treats that record as a curriculum. It looks across finished files for the patterns a busy team never has time to see — the facility that always rejects the first packet, the verification source that's reliably slow, the correction reviewers keep making — and turns each one into a concrete proposal: a new rule, a sharper prompt, an added checkpoint, a fresh integration to hand the Solutions Architect. Nothing it proposes goes live on its own. Improvements are drafted with their evidence and expected effect, tested against the eval suite, and put in front of a human to approve, adjust, or reject. The result is a system that quietly gets better every week — and a team that decides, every time, what 'better' means.
- ReflectsReads completed operations for recurring friction, near-misses, and slow steps.
- ProposesDrafts new skills, rules, and playbooks — as suggestions, never live changes.
- Human-gatedEvery improvement is reviewed and approved by your team before it takes effect.
Questions
Does the Evolution Agent change how the system behaves on its own?
No. It only proposes. Every suggestion — a new rule, a prompt change, an added checkpoint, a new integration — arrives with its evidence and expected effect, is validated against the eval suite, and waits for a human to approve it before anything changes in production.
What does it actually learn from?
The full record of completed operations: requirement outcomes, tool calls and their timing, escalations, and the corrections reviewers make. It looks for recurring friction — a facility that always bounces the first packet, a source that's reliably slow, a mistake that keeps needing the same fix — and proposes a durable improvement for each.
How is this different from the Review Agent?
The Review Agent audits work in flight — catching weak evidence on today's files. The Evolution Agent works across finished files over time, improving the system itself so the same problems stop recurring. One guards quality now; the other raises the ceiling.
See it work a real file.
Thirty minutes, one placement, worked live — start to submit-ready.