A loop that runs for weeks accumulates evidence about itself. Doing nothing with it wastes it; acting freely on it is how an autonomous system drifts from what its owners agreed to, one defensible step at a time. So the loop may change exactly three things.

1. Promote a learning

memory.autoPromote.enabled. A lesson seen recurrence.minSightings times is promoted by the retro itself, attributed to loop-auto, at most recurrence.maxPerRun per run. See memory.

2. Move a knob

tuning:
  enabled: true
  maxChangesPerRetro: 1
  commit: false
  knobs:
    - path: delivery.review.minSeverity
      metric: review-findings-ratio
      values: [nit, med, high, blocker]
    - path: delivery.maxFixRounds
      metric: fix-rounds-per-merge
      min: 1
      max: 4
      step: 1

A knob is auto-adjustable only when it declares the metric that justifies moving it — and that same metric is what moves it back. The next retro compares: worse than before the change means the knob returns and is frozen until a human clears it. A metric at zero never moves a knob, because a loop that keeps tightening a healthy gate eventually stops merging anything.

The edit is written into loop.config.yaml in place, with every comment preserved, validated before it is kept, and recorded in <stateDir>/tuning.json with the reason and the evidence.

Never auto-adjustable, whatever the config says: models, providers, gates and branches — the things that decide who pays and what reaches production.

3. Improve an installed agent

agents.autoImprove. The retro finds the role with the worst outcome-per-run ratio and proposes one dated note appended to that agent's instructions — not a diff written by a model.

agents:
  registryPath: agents.registry.yaml
  autoImprove: true
  evalCommand: [pnpm, run, eval]
  maxAutoLines: 5

The eval is the gate, and the bar is not worse, not better: a change that leaves the score alone but makes the agent clearer is worth keeping, and a change that cannot be measured is never adopted — with no evalCommand, nothing is ever adopted. A failing eval puts the file back exactly as it was, because an agent left half-improved is worse than one never touched.

Three things the machine never does here:

• touch a critical role (architect, reviewer);
• write more than maxAutoLines lines;
• publish anything back to the registry.

And one it stopped doing in 0.15.0: writing to an agent that is code. npx agentskit add <id> installs an agent as a directory of code; an HTML comment appended to agent.ts is a syntax error, not an improvement. The harness now edits only a markdown instructions file that already exists, and records everything else as needs-human with the note and the evidence intact. loop doctor's agents.registry check says which instructions file it found for each installed agent.