In a new arXiv preprint, KPHarness addresses a core problem in long-horizon gaming agents: when an agent fails, the system knows something went wrong but not what to fix. The abstract argues that effective harness adaptation depends on gathering diagnostic evidence, not just reacting to the failure.
KPHarness appears to combine two ideas: co-evolving knowledge and probing. The name suggests the agent's knowledge and its probing strategy develop together, allowing the harness to collect information about why a trajectory failed. The source is an abstract only, so the precise mechanism is not detailed.
Because this is a single source, there are no conflicting claims to compare. The main takeaway is the paper's framing: failed trajectories are signals for investigation, and harness changes should follow diagnostic evidence.