AI 工程从零到一（中文）
Given a target app (web or desktop) and a list of tasks with gold trajectories, build an eval harness.
Produce:
- Task definitions:
(tid, description, gold_steps, success_predicate, state_reset). - Runner: runs the agent, captures every action, records step count + elapsed time + success state.
- Trajectory-efficiency metric:
agent_steps / gold_steps. Report per-task and aggregate. - State reset between tasks — never run one task on state dirtied by another.
- Failure-mode classifier: for each failure, tag whether it's a grounding miss (wrong element) or a planning miss (wrong action).
Hard rejects:
- No state reset between tasks. Cross-task contamination invalidates all scores.
- Success-rate-only reporting. Trajectory efficiency is the 2026 standard.
- Screenshots-only harness without DOM parity. Some agents use DOM+vision; give both unless specifically constraining the surface.
Refusal rules:
- If the tasks have no gold trajectories, refuse. You cannot measure efficiency without them.
- If the app is not pinned to a specific version, refuse. Drift invalidates cross-run comparisons.
- If the agent has destructive tools (delete, publish), require a sandbox copy of the app.
Output: tasks.py, runner.py, failure_classifier.py, report.py, README.md explaining reset policy, gold-trajectory sourcing, and the grounding-vs-planning split. End with "what to read next" pointing to Lesson 21 (computer use models) or Lesson 30 (eval-driven development).