Harness Engineer
The model is a component. The harness is the product.
For teams whose AI demo works and whose production system does not. This seat builds the evals, tool contracts, traces, and failure budgets around the model.
It works in the demo. I need someone who can make it pass evals.
What the Harness Engineer owns
Task suites
Graders that define good before anyone ships.
Tool contracts
Sandboxes, retries, and traces for agents.
A failure budget
What the system may get wrong, and what stops it.
This is a person on a fixed milestone. Eval products are tools a team operates. They are not a substitute for this seat. The alternative is a hire, this seat, or nobody on call.
Where this seat sits
AI Pivot Squad
Primary home. The squad ships the system. This seat makes the model pass evals.
Business Process AI
Evals on workflow agents, when the thing in production is a process rather than a product feature.
When a generated repo is the mess, this seat is the stay-behind after the audit. That audit is a separate quote.
Your users are already on the demo. Make that path hold.
For startups whose proof of concept now has users. The primary seat is a Harness Engineer. AI Pivot Squad hardens the path those users already hit. The leave-behind is a harness and a named owner, so the next release does not depend on the founder watching it.
It worked as a POC. Now people depend on it, and we need to get to the next level.
"Next level" means the path passes evals on real tasks, the cost is bounded, and someone besides the founder can change it. A fundraising narrative is not the deliverable.
The harness is the product.
Task suites, a failure budget, traces, and a cost ceiling. The model stays a component.
Repair the path users hit.
Delete what nobody owns when the repo was generated. Leave the rest of the tree alone.
The first hardened path is a four-week milestone.
A second person on your side is trained on that path. If there is no one to train, the stay-behind is this seat on a separate milestone.
- Cannot yet say what "good" is: an AI Product Manager for a discovery week, then engineering.
- Copilot, Cursor, or an agent wrote the tree and nobody will own it: AI Code Cleanup first.
- It is live and spend or behavior is moving: AI Ops.
- The opening move is this seat and a Squad harden. It is not Business Process AI. It is not a new platform.
Common questions
No. The model is a replaceable component. The harness is the product. A generalist who knows a framework is a different hire.
No. Products are tools a team operates. This engagement is a named person and a fixed milestone.
Yes. It is scoped separately from the $10,000 audit. The audit buys the scored report. The stay-behind is the next quote.
Make the demo pass evals.
A 30-minute scoping call is enough to tell whether this seat is the entry, or whether the squad should come with it.