Building a Harness

Assemble a working agent harness from scratch in Python: the loop, a versioned system prompt, a tool surface, a sandbox and permission modes, hooks, subagents, compaction, and the traces that let you debug it.

10 lessons · about 110 minutes

What you'll go through

• 01The part nobody publishes9 min
• 02The loop is fifteen lines, and you own every one of them10 min
• 03The system prompt is source code that nobody put under review11 min
• 04A tool definition is an API whose only consumer reads under pressure11 min
• 05A read tool that hides the path makes precise edits impossible11 min
• 06The model never executes anything, which means there is exactly one place to stop it12 min
• 07A permission prompt everyone approves is worse than no prompt at all12 min
• 08Delegation buys you a clean window, and parallelism is the tip11 min
• 09What survives turn sixty11 min
• 10Reading a harness you did not write12 min