Agent Harness

The model chooses an action; the surrounding software makes that action possible and returns its result. Anthropic separates its harness from the session log and the sandbox that executes code. Other implementations draw the boundary differently, but the harness coordinates the model’s repeated interaction with those services.

In episode 69 at 15:29, Joel Hron describes giving Thomson Reuters’ model access to Westlaw and Practical Law through a simple agentic harness during training. The harness supplies access; training changes how the model uses it.

For example, if an AI agent chooses the right tool but gets an authorization error, inspect the connection and permissions. If working access still leads to repeated wrong tool choices, inspect the instructions, context and model decisions. Traces of requests, tool calls and results help locate the failure before a team changes the wrong component.