M1 Enforced harness

Hooks, permissions, subagents and durable sessions that make an agent harness enforce rather than suggest.

Lessons

• 01 A harness is a loop plus a policy
• 02 Hooks, permissions and modes
• 03 Subagents
• 04 Durable sessions and idempotency
• 05 Context engineering for code tasks
• 06 Failure taxonomy and caps
• 07 Interview angle

Exercise

M1 enforced harness lab: build your own enforced harness around the Agent SDK, with tests and hints.

1. 01-policy-engine: a Guard that decides what a tool call may do under writable roots and blocked command patterns.
2. 02-hook-adapter: the guard wired as a PreToolUse hook that denies on any doubt, including its own errors.
3. 03-session-store: durable ticket sessions with a fresh-session fallback when a resume fails.
4. 04-idempotent-tools: side effects that run once per key, across restarts.
5. 05-caps-and-report: turn and budget caps and a run summary that says why the run stopped.
6. 06-subagent-verdict: a read-only reviewer subagent whose verdict is validated against a schema.
7. 07-failure-taxonomy: classify harness failures from recorded traces (judgment stage).
8. 08-live-ticket: one tinyshop ticket end to end against the real Agent SDK (needs a Claude login).