Harness guide
What is an agent harness?
A harness is the runtime and safety layer around an agent. It decides what the agent can see, which tools it can use, how actions are verified, and when a human must approve the next step.
Why harnesses matter
Powerful models are useful only when the surrounding system gives them the right job, the right context, and the right guardrails. A harness turns a model from a free-form answer generator into a bounded worker with logs, permissions, and verification.
More autonomy without a harness usually creates more risk. A good harness narrows the action space, records what happened, and makes review possible.
What a harness controls
Inputs
Which docs, files, memories, environment facts, and user requests enter the agent's context.
Tools
Whether the agent can search, edit, run commands, call APIs, or report progress back to a coordinator.
Permissions
Which operations require approval, which paths are off limits, and which secrets must never enter the model context.
Verification
How the harness runs tests, checks diffs, captures failures, and decides whether work is complete.
Local runner harness in Amistio
Amistio's local runner is a harness around approved project work. It runs on the user's machine, inside the paired checkout, with the user's local AI tool setup. The web app tracks state and approval, while the runner provides the controlled execution boundary.
- The web app queues wishes, generated brain artifacts, approvals, and runner-visible work.
- The runner claims only approved work for the paired repository and reports status back.
- Repository source, local credentials, command lines, and arbitrary local paths are not uploaded as public status data.