Agent and harness benchmarks

Harnesses that run an agent, or a whole CLI coding agent, on tasks in a sandbox and grade the outcome, usually with tests the agent cannot see; use them to compare models, harnesses, and configurations on the same tasks.