Process
Start slow
You might say: canary rollout, throttled start
Start with one useful task and one automated trigger at a time, throttled to a small event sample or a once-daily clock. Require human review on every pull request and cap rework at a few tries. This is level 3, where a person reads every diff. Add triggers and open the throttle only after checks and review evidence earn trust; the trust loop then moves the factory toward level 4.
- What it does
- Exposes weak instructions, missing checks, and unsafe permissions while the blast radius is small.
- Ask
- What is the smallest live loop that can prove value and reveal failure?
What to look for
- A canary label or sampled trigger rule
- A low concurrency setting and human approval on every pull request
- A change log showing one trigger or gate added at a time
Sources
- Anthropic, Building effective AI agents, When (and when not to) use agents
we recommend finding the simplest solution possible, and only increasing complexity when needed.
- GitHub, Best practices for using GitHub Copilot to work on tasks, Choosing the right type of tasks to give to Copilot
Initially, you might want to start by giving Copilot simpler tasks, to see how it works as a cloud agent.
- Martin Fowler, Refactoring: Improving the Design of Existing Code, Refactoring
By doing them in small steps you reduce the risk of introducing errors.
- Anthropic, Building effective AI agents, Agents
We recommend extensive testing in sandboxed environments, along with the appropriate guardrails.
- Anthropic, Building effective AI agents, Agents
Agents can then pause for human feedback at checkpoints or when encountering blockers.
- Anthropic, Building effective AI agents, Agents
it’s also common to include stopping conditions (such as a maximum number of iterations) to maintain control.
- GitHub, About protected branches, Require pull request reviews before merging
can require that all pull requests receive a specific number of approving reviews before someone merges the pull request into a protected branch.
- Factory, Software Factory documentation, Filter which deliveries start a run
starts a run only for payloads that carry that field, which keeps a busy event stream from running the automation on every message.
- Factory, Software Factory documentation, Review deliveries
The automation already started 10 runs this hour. The limit resets at the top of the next hour.
- Factory, Software Factory documentation, Schedules
A frequency without a time, such as daily, runs at 9:00 UTC.
- GitHub, Best practices for using GitHub Copilot to work on tasks, Adding custom instructions to your repository
If 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.
- Lauren Tan, pstack, Earn the trust before the loop
Make the loop autonomous only after all four hold. Until then, run it while you watch.
- Matt Pocock, 11 Tips For AI Coding With Ralph Wiggum, 2. Start With HITL, Then Go AFK
You'll understand what it does, refine your prompt, and build confidence before going hands-off.
- Matt Pocock, software-factory skill (pull request #1316), Start slow
Add one automated trigger at a time, throttled at first (one bug report in ten, a clock once a day), with human review on every PR.
- Lauren Tan, Noodle, docs/reference/configuration.md
Start with max_concurrency = 1 or 2 while you’re learning the system, then scale up once you’ve seen the cost per session on your workload.
- Dan Shapiro, The Five Levels: from Spicy Autocomplete to the Dark Factory, Level 3
You are the human in the loop. Your coding agent is always running multiple tabs. You spend your days reviewing code.
Use it on your project
teardown-apply: hold "Start slow" (start-slow) 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