What the Harness Costs
Explore how to measure and attribute the costs of controls within a coding agent's harness. Learn to analyze trade-offs between runtime, human attention, and maintenance across multiple sessions. Understand strategies to optimize verification processes by scoping controls properly and pricing them against the failures they prevent, helping you design more cost-effective, reliable coding agents.
We'll cover the following...
- What did the finished ticket actually cost?
- Why is cost per session the wrong number?
- Which budgets does a harness spend?
- How do you attribute cost to one control?
- What does an expensive control look like?
- How should the price change what the loop does?
- Which controls should pay for themselves?
- When is a control worth keeping at a loss?
- Can you price Wrenfold’s stack?
- What can you add to your own harness today?
- What’s next?
Wrenfold’s unattended loop finished CLIN-547 on a Thursday afternoon. Waitlist promotion worked, the retry was idempotent, both receipts joined on the same revision, and the exit gate passed.
The run took five hours and forty minutes across fourteen attempts.
Trace review showed why: every attempt reloaded nine contract files, re-derived the stored Claim-to-Proof verdict, and ran a full database reset between attempts.
The contracts were correct and had each earned a place. They had never been priced.
What did the finished ticket actually cost?
Almost two of those hours bought nothing that changed an outcome.
The reseed was the largest single line item: fourteen full reseeds at just under four minutes each, where a scoped reseed of one provider’s appointments would have taken seconds and protected the same invariant. The contract reloads came second. The matrix re-derivation came third, and it produced a result identical to the one already stored.
None of that is a design flaw in the individual contracts. It is the predictable result of adding controls one at a time and never asking what the accumulated stack spends per unit of work delivered.
Anthropic’s harness work makes the same point sharply. Comparing a solo agent against a full harness on the same task, the solo run took twenty minutes and nine dollars and produced a game where “my entities appeared on screen but nothing responded to input.” The harnessed run took six hours and two hundred dollars, and the game worked. Over twenty times the cost bought the difference between a demo and a deliverable.
That trade was clearly worth making. Whether every component inside it was worth making is a separate question, and it needs a number.