# Intro to the Harness

An agent harness is the layer around an agent that gives you control, structure, and visibility.

It usually handles four jobs:

• Containment: keeps the agent within a scoped environment
• Inputs: supplies prompts, files, tools, and context
• Execution: runs the agent and captures what it does
• Observability: records logs, traces, outputs, and decisions

A harness is not the agent itself. It is the system that makes the agent safer to run, easier to inspect, and more reliable to repeat.

Example

If you ask an agent to review a document, the harness might:

1. Load the document
2. Provide a limited set of tools
3. Record each tool call
4. Capture the final answer
5. Save logs for later review

Without a harness, the agent may still answer. But you lose control over where it reads, what it can touch, and how you debug mistakes.

Key ideas

• A harness wraps the agent
• Containment limits what the agent can access
• Observability helps you see what happened
• A good harness makes agent behavior repeatable

Micro task

Look at one agent workflow you use or imagine. Name one thing the harness controls and one thing it records.