Arca runs a nightly meta-agent across real-world agent execution traces and staff feedback to identify harness errors, misleading tools, and missing knowledge