Harness and Cosmos: the factory keeps owning a service after the merge, and auto-approval widens one change class at a time

Harness's write-up of how it fits together with Cosmos, the agent platform it got by acquiring Augment Code, describes factories that run continuously, with people stepping in only where their judgment is needed. Work starts from events, a PR opening, an alert firing, or a ticket being filed, and the factories themselves are defined as code, so they are versioned, reviewed, and rolled back like any other change.

The PR is not the finish line

The new idea in this post is about what happens after merge. The factory is not finished when the pull request opens: it is assigned to a repository, a service or a class of finding, the way a team is. If verification fails, the same Experts see the evidence. If a revert fires, they pick the work back up. The authors contrast this with a coding agent that opens a PR and disappears. Production signals become new work, and review comments, incident follow-ups and triage corrections are stored in Expert memory, so each change starts with what the last one taught.

Autonomy is earned from the record

Auto-approval is narrow by design. cosmos approve approves a PR only when the requester owns every changed file, no review findings are open on the current commit, and the Verifier has found no contradicting evidence, and a human still makes the merge. On the Harness side, a worker agent's token is minted per run, scoped to the intersection of the agent's permissions and the triggering user's RBAC, and the same policies that gate production deployments gate the agent, so it cannot merge to main even if its prompt tells it to.

The widening is gradual. Verification results, revert rates, and audit trails show which kinds of change the factory handles reliably, and organizations use that record to widen auto-approval one class of change at a time. The authors' framing: governance is what lets an organization give agents more autonomy. They also warn that adoption lags deployment, because teams needed weeks to tune guardrails and learn to trust each agent.

The numbers, and whose they are

On the Cosmos team's own development, after review moved to Cosmos, weekly output tripled while bug-introducing commits per unit of output fell from 0.097 to 0.006. Its Incident Investigator handles 81% of incidents and cuts the median time to first RCA from 30 to 6 minutes. These are a vendor's figures about its own product. The merge-time and revert numbers repeated in the post were covered in Augment's factory case study.

For teams running agents around the clock, the useful part is the shape of the loop: the agent stays responsible for its change after merge, failures come back to it with evidence attached, and the scope of what merges without a close look grows only as fast as the track record supports. Compare the earlier note on how Cosmos specializes the PR-to-merge loop.