What's this about?
HarnessDev (226 upvotes) shifts evaluation from task outputs to runnable infrastructure: the agent starts from a minimal seed, builds a complete execution system, then iteratively revises its own harness using downstream feedback.
What exactly is being tested?
Not "can you solve the task" but "can you build and improve the machine that solves tasks" — self-modifying scaffolding, one level up from self-modifying answers.
Why is this worth noting?
If harness quality drives capability as much as this week's evidence suggests (ARC's 37-point gap, last week's 17× cost spread), then agents that improve their own harness compound. It also puts capability research and security research on a collision course — the hook-attack paper is about exactly this layer.