Loop

Improvement loop

You might say: retro loop, the factory learns, friction loop

Finished runs leave logs and review findings. A retro, by a human or a retro run, finds a mistake that repeats and changes the factory's instructions or checks. A check holds better than a written rule, because nothing fails when an agent skips a rule, so each repeated review comment should become a check where it can. Later runs make that mistake less often, so more of their pull requests merge without a person.

What it does
Turns each repeated mistake into a lasting fix in the factory, so quality grows with use.
Ask
Which repeated failure should change the factory next, and which measure will show the change worked?
Trouble
The same review comments appear week after week, and nobody edits the instructions or adds a check.

Loop: reinforcing

Run log → Instructions → Ratchet → Merge rate ↺

Sources

  1. Anthropic, Building effective AI agents, Combining and customizing these patternsthe key to success, as with any LLM features, is measuring performance and iterating on implementations.
  2. Donella Meadows, Leverage Points: Places to Intervene in a System, 7. The gain around driving positive feedback loopsPositive feedback loops are sources of growth, explosion, erosion, and collapse in systems.
  3. Betsy Beyer et al., Site Reliability Engineering, Chapter 1, IntroductionThis investigation should establish what happened in detail, find all root causes of the event, and assign actions to correct the problem or improve how it is addressed next time.
  4. Betsy Beyer et al., The Site Reliability Workbook, Chapter 2 - Implementing SLOsYour first attempt at an SLI and SLO doesn’t have to be correct; the most important goal is to get something in place and measured, and to set up a feedback loop so you can improve.
  5. Gene Kim et al., The DevOps Handbook, Part I - IntroductionThe principles of Continual Learning and Experimentation, which foster a high-trust culture and a scientific approach to organizational improvement and risk-taking as part of our daily work.
  6. GitHub, Best practices for using GitHub Copilot to work on tasks, Adding custom instructions to your repositoryIf Copilot is able to build, test and validate its changes in its own development environment, it is more likely to produce good pull requests which can be merged quickly.
  7. Lauren Tan, pstack, Fix the environment with /correctWhen you correct agents for the same mistake again and again, the fix belongs in the repo, not in your next prompt.
  8. Lauren Tan, pstack, Fix the environment with /correctWrite it down as a doc or agent rule. Nothing fails when an agent skips a rule, so this comes last.
  9. Kody, kody-hq/kody repository, docs/contributing/harness-engineering.mdWhen a mistake repeats, encode it in the strongest cheap guardrail that actually works. Prefer a failing check over another paragraph of advice.
  10. Kody, kody-hq/kody repository, docs/contributing/harness-engineering.mdRule of thumb: if reviewers repeat the same comment twice, encode it.
  11. Kent C. Dodds, kcd-skills, autopilot skillAgents file friction after every PR with a priority; other agents fix it, highest priority first.

Use it on your project

teardown-apply: hold "Improvement loop" (improvement-loop) from the teardown at https://teardownatlas.com/software-factory/ against my project

Paste it into a coding agent that has the teardown skills.

An AI agent wrote this map with the teardown skill. It can contain errors: check the sources before you rely on a part. About and legal