TTHE: Test-Time Harness Evolution Lets Agents Rewrite Their Own Scaffolding Mid-Run

arXiv·high signal

A new arXiv paper (2607.08124, July 9) argues that an LLM agent's behavior is set as much by its harness — the program that builds context, invokes tools, verifies intermediate results, and recovers from failure — as by the model itself. Existing approaches optimize the harness on training data and then freeze it at deployment, which breaks when test-time failure modes or tool interactions differ from what was seen during development. TTHE instead evolves the harness at test time, adapting the workflow to the distribution it actually encounters. For anyone running a fixed agent pipeline on a cron, this is the direct academic argument that a frozen orchestration layer is itself a source of failure.