AutoReview Crew

Four AI agents review a code change together in Band, recruit a test specialist mid-review, and escalate risky changes to a human before merge.

Recorded take: June 15, 2026, 13:02 UTC

This replay mirrors the real Band transcript used for the demo video — expand any line to read the exact message the agent posted in the room. The live system runs through run_crew.py; this hosted page is a safe no-key companion.

About 51 seconds in the real Band room.
Decision: ESCALATE_TO_HUMAN. All specialist reports arrived before the Lead finalized.

The agents

What the crew caught

How the system is wired

Band is the coordination layer

Assignments, reports, recruitment, shared context, and human escalation all happen through Band room messages and platform tools.

Deterministic Lead coordination

The Lead waits for current-round reports, ignores stale context, recruits Test only when needed, and escalates risky changes.

Cross-provider reviewers

AI/ML API powers the Lead, Correctness, and Test agents; Featherless powers the Security Reviewer with Qwen2.5-72B.

Human-in-the-loop merge gate

The crew does not approve risky code silently. Critical findings are aggregated and sent back to the human owner.