- Add types, tests, and CI before adding agents
- Review architecture and test quality, not every line
- Give agents their own branches and CI runs
- Learn agent management as a deliberate skill
- Start now
I used to trust my compiler to emit correct output and never read the bytecode. If something was wrong, the bug was in my source. Nothing has changed: I now write specs, and the LLM compiles them into code.
Well written specs next to your code are vital for getting a good output from an agent. Together with quality gates like tests, linters and etc you can get great output from your agent. Managing your back-pressure is the key to running an autonomous agent
When the agent runs a quality gate command, test, linter, etc, you don't need to feed the success output back into the agents' context. With agent hooks or in the case of #Elixir with a custom mix task, you can have succesful commands like mix test simply return checkmark rather than 5k dots.
Frontline LLM's are capable of following ~150-200 instructions. Before CLAUDE.md is read, claude already gets ~50 instructions. This means that the amount of instructions in your personal CLAUDE.md file plus any in your project, ideally need to stay below ~150.