Chapter 9: Wire Up a Complete Harness
Every harness starts here: a model sitting idle. It can generate text but can't do anything with it. No tools. No loop. No memory. Just raw intelligence waiting for infrastructure.
Wrap the model in Think-Act-Observe-Decide. Now it can iterate, but notice the "Act" step has nowhere to go. The loop runs but nothing happens. A loop without tools is like a car engine without wheels.
Three tools: Read, Bash, Search. Now the loop's "Act" step can actually touch the world. The dot travels from the loop to tools and back. The agent can read files, run commands, search the web.
Diamond gates on every wire. Read is auto-approved (green). Bash requires confirmation (yellow). Search is auto-approved. Now the agent can't execute dangerous commands without your permission.
A context bucket to manage the token window. A session log to survive crashes. The agent now knows what it's done, what it needs to do, and can resume from where it left off.
Watch: Read fails with an error. The harness injects the error into context, and the model switches to Search. Green checkmark, recovered. This is a complete, production-ready harness.
Model + Loop + Tools + Gates + Context + Memory + Errors = Agent.