A weekly gardener turns the fixes your team accepted into lint rules, so Monday's correction reaches Thursday's agent
Oct 19, Jag Reehal. Shepherd fixes a PR inside the coding agent, Stamp approves or refuses it in CI, and once a week a gardener reads the GitHub record and proposes one small change to skills or lint rules, as a PR with evidence a person merges. It counts only unreverted fixes, needs three accepted cases before proposing a rule, and its first run over fourteen PRs produced one edit.
Source: Your coding agents need a gardener
Jag Reehal starts from a familiar waste: "You correct an agent on Monday. On Thursday a fresh session makes the same mistake, because nothing it reads has changed." His fix is a third loop on top of review. The number he optimises for is how much of your attention the agents use.
Three parts, each with one kind of authority
Shepherd runs in the coding agent, reviews a PR, and fixes what it can, including the build. Stamp runs in CI and approves or refuses; it never changes code. The gardener runs once a week, reads the GitHub record (review threads, Shepherd's commits, reverts, statuses, round counts), and proposes changes to Shepherd's skills or the repo's lint rules as a PR that a person decides on.
The evidence bar is strict. A fix counts only when Shepherd named the rule in a commit trailer, nobody reverted the commit, and the PR came from someone who can push to the repository. A resolved thread counts for nothing, because triage also resolves findings it judges mistaken. A lint rule needs the same finding fixed on at least three PRs and ships as a warning, with the cases it must flag and a counterexample it must leave alone. A regression that Stamp approved can justify a check from a single incident, because it comes with a reproducing test.
Small power, small output
The agent that reads PR text has read-only tokens and writes to an outbox; a separate job on a fresh runner applies patches, refuses changes to .github and .git, and opens the PR. Edits to Stamp's own rules go to a person. The first run, over fourteen PRs on his test repo, dropped most signals as already covered and proposed one edit. He wanted a gardener that proposes little with evidence over one that rewrites prompts every week. What he hasn't built yet is a set of representative tasks to rerun against a revised skill.
This is the slow, compounding side of a factory: each week's accepted corrections become checks the next run can't skip, without anyone driving a session. Compare Lauren Tan's gardener sampling merged work at 2,500 PRs a month and yureki_lab's replay-gated rule loop for a 24/7 agent.