Engineering blog Post 06
KrewHQ: running an engineering programme with a crew of AI agents
One day a release commit quietly rewrote evidence bundles that were already on main, so each manifest claimed a run at its original date with a lockfile that did not exist then. The author was an AI coding agent.
The review board, an independent AI reviewer agent, called it false provenance and ruled the release a no-go, and every file was restored byte for byte the same day. This is how one engineer ran a crew of AI agents, and why that review board existed.
Built beside Kiro Crew
KrewHQ runs on Kiro Crew. Each team works in its own Kiro Crew instance, a spoke, and the hub runs its own instance too. KrewHQ is the layer between them: shared, scoped, auditable team memory beside Kiro Crew, with a hub API and a spoke for each team.
There are six spokes: the SDK, the SDK harness, ConstellOps, the ConstellOps harness, the robots fork and the programme lead. KrewHQ is the owner's own product, built beside Kiro Crew. Kiro Crew is named here only as what it runs on; no endorsement is implied.
The team map
The map is one team per repository, plus a programme commons. These are the comments at the top of the map file:
# One team per repo. Knowledge crosses teams only by explicit promote / handoff (scoped + provenance).
# `program` is the shared commons: every team is a member, so records promoted or handed off
# under program scope are readable by all teams. Team-scoped records stay private to their team.
Each team has a lead agent, with helper agents inside the team. The owner is the engineering manager of every team. A team reads and writes only its own repository and the shared docs, and uses other teams' work only as tagged releases.
Handoffs: a link, not a copy
The most important loop runs between a harness and its product. The harness files a finding with a case, a reproduce command and evidence. The product team fixes it and cuts a patch tag.
The fix is handed off through KrewHQ as a link to that tag, with one line of state and next step. The harness re-runs the case against the tag, and marks it verified or reopens it.
Handing off a link rather than a copy keeps one source of truth. Verification is event-driven: harness leads watch for handoffs and start the re-run without being asked.
An independent review board
The review board is a fresh AI reviewer agent, never the author of the change, working in a clean context. It sees the pull request, the programme rules, the metric definitions and the design.
It must approve changes to the gates, the uplink path, autonomy levels and public interfaces, plus every release.
Its checklist includes: gates fail closed and cannot be bypassed; no fabricated numbers; no secrets or account identifiers; teams stay isolated. Verdicts are approve, request changes or reject, with conditions, and each is logged.
What "independent" means here
Every reviewer in this programme is an AI agent: independent reviewer agents with no shared context. No outside human has reviewed the code yet.
It is the same kind of model as the authors, so its independence is a clean context and no stake in the change.
It is not a different mind, and the owner's rejections (post 8) are partly why we know that matters.
Evidence gates before merge. Hosted CI was off on the private repositories. Merging needs board approval at the exact head commit, plus a recorded local verification at that same commit. If the head moves, it is verified again.
What went wrong
Re-records under one rule
During a live recording campaign, the network on the recording host failed twice. Runs recorded during the outages were unreliable, and the temptation is to pick the bad-looking ones by eye.
Instead, one criterion was written down and applied to every run: logged agent errors and failed model streams in that run's journal, or zero model calls.
Runs that matched were set aside with a note, their spend stayed in the accounts, and they were re-recorded. The board re-applied the criterion to every run folder and found it selected exactly the set-aside runs, with no mismatches.
The same board ruled that a run which failed to reproduce for a different reason could not be re-recorded away. Post 4 has that story.
The key-person question
The plain answer first: the founder, Vivek Raja, is the only human on this programme, and no outside human engineer has reviewed the code yet. Everything above reduces the risk; none of it removes it.
What does exist is a written trail a new engineer can read: every decision, review verdict and handoff is a record with scope and provenance, and every release has an evidence bundle.
An outside human code review is the next step we want, before any operator pilot.
Next pass
A crew of agents can build fast, and it can also rewrite its own evidence. What made this programme trustworthy was not the agents; it was the board that could say no-go, and the rule that recorded evidence is never rewritten.