AI 工程从零到一（中文）

Given a task description, pick Crew (autonomous) or Flow (deterministic), then scaffold.

Decision:

1. Does the task have SLA, compliance, or deterministic replay requirements? -> Flow.
2. Is the task exploratory (research, first draft, brainstorm)? -> Crew.
3. Does the task have 4+ specialists with LLM-picked ordering? -> Hierarchical Crew.
4. Does the task have <=3 specialists in a fixed order? -> Sequential Crew or Flow — prefer Flow.

For Crews, produce:

1. Agent definitions: role, goal, backstory (tight, <=200 words), tools.
2. Task definitions: description, expected_output, agent.
3. Crew with the right Process (Sequential | Hierarchical).
4. A test harness that runs the Crew on sample inputs and checks that expected_outputs are produced.

For Flows, produce:

1. @start entry function.
2. @listen(topic) steps forming a DAG.
3. Explicit event topics; no magical broadcast.
4. A replay harness: given a kickoff payload, rerun deterministically.

Hard rejects:

• Crews without backstories. Backstories are load-bearing.
• Flows without explicit topic names. "Implicit chaining" defeats the audit purpose.
• Hierarchical Crews with 2 specialists. The manager overhead is not earning cost.

Refusal rules:

• If the user asks for a Crew on a prod-only compliance task, refuse and migrate to Flow.
• If the user asks for a Flow on an open-ended research task, refuse and migrate to Crew.
• If the backstory exceeds 200 words, refuse and require a trim. Context budget is finite.

Output: agents.py, tasks.py, crew.py or flow.py, plus README.md with the decision rationale. End with "what to read next" pointing to Lesson 24 (Langfuse/AgentOps) for observability, or Lesson 13 if the Flow needs durable resume semantics.