Scope the sentence, attach sources, run checks, and block language that does not survive.
Test the workflow where it actually runs.
A harness wraps a workflow, system integration, tools, state, permissions, memory, and validation loops. The point is not autonomy theater. The point is a compliance-ready run: what it did, what failed, and what a human still has to approve.
Harnessable surfaces.
Wrap prompts, evals, thresholds, run records, and deployment context around behavior.
Bind handoffs, queues, owners, approvals, and receipts into a compliance-ready operating loop.
Pin sources, transforms, lineage, clocks, and readouts so the row can defend itself.
Give a base model tools, state, memory, permissions, tests, and a receipt-bound loop.
Control APIs, auth, retries, schemas, health checks, and failure modes before scale.
Harness architecture.
Defines the boundary, interface, allowed inputs, and expected outputs for the object.
Sets operating mode, permissions, thresholds, owner rules, and resource limits.
Connects the object to the systems it may inspect, mutate, query, or test.
Preserves goal, context, open changes, decisions, blockers, and run status.
Keeps stable project and domain facts from being rediscovered every cycle.
Runs checks, reads failures, narrows the claim, and emits receipts before handoff.
Work loop.
Intake
Bind the user request to scope, files, constraints, and a declared success gate.
Act
Let the agent use approved tools to inspect, edit, run, and refine the code.
Validate
Run tests, linters, browser checks, fixtures, or probes that can actually fail.
Receipt
Return the diff, the checks, the failures, and the next compliance or approval gate.
Readout states.
Engineering handoff.
The harness knows which object is active and what changed.
Stable customer, domain, and repo facts survive across the work loop.
Every action runs through named tools, permissions, and boundaries.
The agent returns checks, failures, source links, and next gates.