A seven-station agent factory on a trading codebase, and the list of things it won't automate
Jonathan Wage maps TradersPost's agent pipeline stage by stage: a gate at every station, risk that agents can raise but never lower, and a person still deciding every merge.
Jonathan Wage at TradersPost describes the agent pipeline his team runs on a production Symfony codebase that executes trades. The post is a map of a longer series: seven stages, each with the failure it guards against and the gate they built. His one-line version is that agents propose, while deterministic gates and accountable humans dispose.
The stations are concrete. Each agent gets its own worktree, container, and state, and can't start implementing until a setup check passes. The always-on instruction file is treated as a budget enforced in CI, which shrank it about 76%. CI runs on request rather than on every edit, and a gate status on the current head goes stale on the next push. Migrations are parsed with a real PostgreSQL grammar, and warnings block until a different human acknowledges them. Review starts with a deterministic gate, then a fresh reviewer per dimension and a quorum. A pull request goes to a person when CI still fails after 5 author pushes or a finding is disputed across 2 full rounds.
The risk rule is the sharpest idea. A deterministic floor comes from which files changed and who owns them, and agent evidence can only raise that risk, never argue it down. Trade execution, auth, billing, migrations, and compliance work always get human review, whatever any agent says.
Why it matters for code review
Most write-ups of always-on factories measure success in merges per night. This one is useful because it publishes the other column: what the team deliberately does not automate. Marking a PR ready, merging, submitting reviews, merging to production, and downgrading risk all stay with people. Bot approval for low-risk changes like docs and tests exists only as a labelled pilot, and repository rules haven't been changed to trust it.
That makes it a fair counterweight to the lights-out stories. Continuous agent work still compounds here, but on a codebase where a miss moves money, the throughput is built so that every station is cheap and mechanical and human attention is spent only where judgment is the product. Wage's advice on order matters too: add watchers and supervisors last, because automation on a weak gate multiplies the weakness.