What's this about?
Agent harnesses determine how a model uses instructions, tools and runtime components — but adapting a harness requires verifying that the change helped, and verification is expensive. Existing propose-and-verify methods score every candidate the same way. This paper proposes behaviour-aware verification: spend the verification budget where behaviour actually differs.
Why is verification the bottleneck?
Because to know whether a harness change helped, you have to run the agent on real tasks, repeatedly, and real tasks take minutes to hours. Proposing a hundred variants is nearly free; testing a hundred variants is the entire cost. Any method that can tell which candidates are behaviourally identical to something already tested cuts that cost directly — the same insight that makes test-suite reduction work in ordinary software.
Why does this matter to you?
- It is the practical complement to JIT-Agent: that paper says evolve the harness, this one says here is how to afford it.
- Behaviour-aware selection generalises to any expensive evaluation loop, including security scanning.
- Third harness-evolution paper this month — the field has clearly decided the harness is where the gains are.