Skip to content

Agent harness

An agent harness manages the model and tool loop for an agent. It supplies context to model calls, exposes permitted tools, and returns results for the surrounding run. A domain-specific harness uses the persona’s definitions and workspace bindings to keep the work within the team’s scope.

A harness takes the selected playbook step, authorized context, model binding, and available tools. The playbook determines the operation and its constraints. The runtime retains the team scope, run state, budgets, and effect policy. An implementation can provide both responsibilities in one process. The draft leaves the interface between them to each implementation.

  1. Assemble the instructions and context for the active step.
  2. Invoke the model chosen for the step’s capabilities.
  3. Validate any proposed tool call against the declared tool and arguments before dispatching through the authorized path.
  4. Add permitted results to the step context and continue within the budget.
  5. Return validated output, a request for review, or a failure to the surrounding run.

The model can choose which changes deserve attention. The harness still validates that every cited meeting comes from the authorized calendar result and that the brief fits the expected output. The sequence below illustrates responsibilities.

Playbook step: summarize meeting changes
runtime → supplies allowed meeting records and output contract
harness → assembles context and invokes the workspace-bound model
harness → validates the summary and meeting references
runtime → records the result for review or an eligible next step

When an agent proposes to publish the brief, hand the proposed action back to the runtime for its policy checks. A separate approval can bind the destination, acting account, and exact content before dispatch. See tools and governed execution.