Labs / harness design / build a safe loop

Agent Harness Builder

Turn a model call into a system: select the tools it can use, define where a human must approve, set a budget, and add the checks that make a run auditable.

Configure the run

Workflow patternModel tierMax tool calls

Tool permissions Search internal knowledge Read business records Write / update records Run code or shell commands

Control plane Human approval for side effects Store trace + tool results Evaluate outcome after run

Generated harness contract

—

Choose a workflow and build its contract.

Key takeaway: the model is one component. The harness owns permissions, approvals, timeouts, retries, traces, and the evaluator that checks what really happened—not what the agent claimed.