Field notes · 2026 · Series 2
What happened when product managers drove the agent
The last series ended on a question: if the context layer is good enough, does the person driving the agent still need to know the codebase best? We ran the experiment. Four weeks, four PMs writing their own A/B test code, a developer approving every change. Six posts, with the numbers.
Where it stands · 02 / 06 published
The saving isn't the 11.6 hours of review. It's what an experiment costs to propose — a five-day, four-person item has to be prioritised; a fifteen-minute review does not.
Experiments stop being expensive enough for the backlog to ration them, and the ones that never got proposed start getting proposed. The same cost that gets absorbed between other work is also the cost that doesn't get counted — that's the advantage and the risk.
The six posts
02 / 06 published- Product managers got 38 code changes through developer review. All of it took 11.6 hours
Median 15 minutes per experiment, against a documented route of four people, four systems and five days minimum. But total developer time was ~21 hours, front-loaded into setup — and the real saving isn’t hours at all, it’s that experiments stop queuing.
- An AI coding agent waited 30 minutes for a developer, then answered itself
Same ambiguity, same guardrails, same model — the only variable was whether a human replied. The second time nobody did, and the agent downgraded its own stop and committed. The fix took five days and touched no application code: no answer is never permission.
Next · The amber gate we removed, and the tripwire nobody called
Forthcoming
Forthcoming
Forthcoming
Coming next
The amber gate — “ask a developer first” — fired about ten times over three weeks and was approved every time. Each check cost a developer five minutes; each wait cost a PM hours, in one case days.
So we removed it for week four, with a tripwire. Week four was the highest-volume, cleanest week of the pilot — and also the week an experiment the gate would have stopped went through, and nobody called the tripwire.
Picks up from Series 1: What 47 people taught us about AI coding agents.