Ticket to merged PR in 1.2 days instead of 4.5 - with 66% merged untouched

RESULTS

Measured over the 12 weeks either side of launch.

Cycle time: 4.5 days -> 1.2
Ticket picked up to change merged. Most of what disappeared was queue time, not typing time.
First review in 41 seconds
Before a human opens the PR, against a 6-hour median for the first human comment.
66% merged untouched
Two thirds of the agent's merged PRs needed no human edit. It is weakest where judgement is needed: mechanical migrations land 96.8%, a ticket carrying a design decision 59.6%.
+74% merged PRs per engineer
3.1 to 5.4 a week, with no change to headcount and no measurable change in revert rate.
A two-year migration shipped in four days
143 files from moment to date-fns. Nine were handed back where no mechanical equivalent existed.

The problem

Nothing was wrong with the engineers. The time was going into waiting.

Two queues, not one
A ticket waited for somebody free to pick it up, then the PR waited for somebody free to review it. Median first review comment landed 6 hours after the PR was raised.
The backlog nobody would ever start
A 143-file migration had been agreed twice and scheduled never, because it was two weeks of work with no visible outcome.
Coverage drifted down
Tests were the first thing dropped under a deadline, and the modules that most needed them were the ones nobody enjoyed working in.

What we shipped

Agents inside the workflow they already had.

Ticket to PR
Reads the ticket and its discussion, plans the change, writes it, runs the suite, opens the PR.
A review on every PR including its own
With a severity and a suggested patch rather than a comment.
Test generation aimed at uncovered branches
Not at paths that already had assertions. Weighted coverage rose from 57.7% to 74.7%.
Migrations across a repository
Handing back the files where no mechanical equivalent exists, with the reason per file.
Guardrails
No merge rights, protected paths read-only, no new dependencies without a ticket, no schema changes inside a feature PR.

The agent has no merge rights and never has - all 375 runs a month end at a pull request waiting for a person. Four months from first call to go-live, the fastest of our engagements because the users were the buyers. The engagement is ongoing.

Their engineers were not slow. Their work sat in queues - waiting for someone free to start a ticket, then waiting for someone free to review it. We embedded agents into the workflow they already had: reading the ticket, opening the PR, reviewing every PR before a human sees it, and filling coverage behind the change. Cycle time fell from 4.5 days to 1.2, and 66% of the agent's merged PRs needed no human edit at all.