Process

Factory design

You might say: operating model, automation policy

The deliberate work of choosing a problem, adding triggers and services in small steps, and deciding where people stay in the loop. It also covers the human role once agents carry most changes, and the evals that show whether a change to the factory helped.

What it does

Turns an automation idea into a small system that earns more autonomy as evidence comes in.

Ask

What is the smallest useful loop, what can stop it, and what evidence earns the next level?

Parts

• Start slow: Start with one useful task and one automated trigger at a time, throttled to a small event sample or a once-daily clock.
• Buy versus build: Use third-party software for proven capabilities when customization matters less; build the task policies, checks, and feedback that differentiate the factory.
• Autonomy level: How much of the work agents do with no person in the loop.
• Human re-entry: The point where a person comes back after a run ends or a gate blocks progress.
• Slop: Fast, plausible output that reaches review without enough evidence or attention to make it trustworthy.
• Human role: The work left for people once agents write and prove most changes.
• Factory evals: A fixed set of past tickets and scenarios that the factory reruns after any change to its models, prompts, tools or gates.
• Trust loop: What human review catches becomes new checks, sharper instructions and stronger verification.
• Improvement loop: Finished runs leave logs and review findings.

Sources

1. Anthropic, Building effective AI agents, When (and when not to) use agentswe recommend finding the simplest solution possible, and only increasing complexity when needed.
2. Gene Kim, Vibe Coding: Pairing vs. Delegation, Vibe Coding: Pairing vs. DelegationYou delegate when you’re sure the task novelty is not terribly high
3. Factory, Software Factory documentation, Custom AutomationsCreate, configure, and monitor automations that run Droid on a schedule, from Slack messages, on GitHub events, or from inbound webhooks.
4. Anthropic, Building effective AI agents, AgentsAgents can then pause for human feedback at checkpoints or when encountering blockers.
5. Matt Pocock, software-factory skill (pull request #1316), Designing a factoryBuy the services and build the tasks: the tasks carry the user's design, the services are generic.
6. Matt Pocock, AI Coding Dictionary, Software factoryDeciding which of those decisions stay human is the main design question.
7. Kent C. Dodds, kcd-skills, autopilot skillOrder: P1 first (checkable), then P2/P3 (trust), then loops.

Use it on your project

teardown-apply: hold "Factory design" (factory-design) 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