A software factory is the floor under fast agent hands — merge stays human

Manish Gupta at BlueLupin (September 10) watched the “just let them ship” demo: first feature before lunch, second quietly contradicts it, week’s end a pile of plausible diffs with no session owning the decision. The agent is a station on the floor, not the factory. Four things exist before an agent opens a client repo in anger: a ruleset in the repo (layers, import direction, contracts — no ruleset, no session); a map so cold sessions do not rediscover the tree; a broker that attaches short-lived credentials after checking agent/session/args (no service account in the prompt); and a human who still reads the diff — the person who opened the session does not approve their own change.

Ticket path is stations, not vibes: intake with out-of-scope, optional short spec, one-task/one-branch/PR then throw the session away, verify with trusted tests plus a break-the-change pass (polite self-review is not a gate), named human, then monitor feeding tickets back. Red CI stops; irreversible actions (merge to default, prod migrations, secret rotation, customer email) never belong to the coding agent. The sharp capacity claim: launching more agents does not raise the merge rate — it fills the review queue. Real capacity is how fast a small squad can define the job, read the trace, and approve a release. Factory definition lives as a PR’d file next to the code; golden tickets replay when the harness changes; start on boring chores, not the feature the client is nervous about.

Why it matters for code review

Harnessed overnight throughput compounds only when the floor rejects bad work before morning triage and merge stays a named human stop. BlueLupin’s stations relocate judgment: machines walk tickets and fail closed; humans keep promote, and FOMO about idle agents is earned only after the review drain has an owner.