3.3.2Build

Building is when humans don't give feedback.

• The goal is to tee up the agent to do a bunch of work while you're away.
  • That means removing the human from the loop wherever possible.
• Let's think through everything you need in place to eliminate the need for human input:
  • First, limit what the agent can do by readying your codebase.
  • Second, settle all the questions about what you want to build upfront during Plan. If you "want to see how it looks" as part of build, then you have to wait for build to give feedback.
  • Third, enable the agent to build the way you want by using skills and rules (See Syllabus)
  • Fourth, help the agent avoid asking you for something it doesn't need to by setting up:
    • Automated secrets management
    • A sandbox environment where the agent has full permissions to do whatever it needs
    • Anything else???
  • Fifth, enable the agent to Verify its own work.

If you've done all this successfully, then:

• Your agent knows exactly what to build
• You've told it what guidelines to follow when building it
• You've given it a way to verify its work

That means that by the time you come in to review the final product, you're going to validate that:

• It works the way you want
• Any new decisions that were made during the build align with your judgement

In theory, you shouldn't need to review any code.

What's the biggest bottleneck?

• I've listed build-readiness steps in chronological order, but in fact you should prioritize based on whatever the biggest bottleneck is. For most people, that's likely verification.