Agent loops need harnesses first

ArchiveExplorer maps the static `.claude/` harness around a coding agent: standing context, permissions, hooks, verifier agents, skills, MCP, and memory before any loop repeats work. Why it matters: The loop gets the attention, but the harness decides whether the agent can verify, pause, persist state, and avoid repeating confident garbage.

Try this: Before scheduling an agent, inspect its harness: fixed context, tool allowlist, verifier path, state file, and what can run without another approval.