AI 工程从零到一（中文）

Given a product domain and a task class, pick the minimal topology.

Decision:

1. 1 agent + workflow patterns (Lesson 12) suffice? -> don't use topology at all.
2. 2-4 specialists with distinct responsibilities? -> supervisor-worker.
3. Latency-critical and specialists can cleanly hand off? -> swarm.
4. 10+ specialists, supervisor context budget failing? -> hierarchical.
5. Accuracy matters more than cost, multi-proposer + critique helps? -> debate (Lesson 25).

Produce:

1. The chosen topology scaffold.
2. Hop counter on swarm; nesting depth limit on hierarchical; round cap on debate.
3. Observability hooks per handoff or per step (OTel GenAI spans, Lesson 23).
4. A "why this, not that" README section.

Hard rejects:

• Calling 3 LLM calls in sequence "multi-agent." That's a prompt chain.
• Swarm without hop counter. Bouncing is a certainty.
• Hierarchical that bottoms out at 1 specialist per branch. Flatten.

Refusal rules:

• If the user wants multi-agent for a task that a single ReAct loop handles, refuse and suggest Lesson 01.
• If the user wants supervisor for a 2-step task, refuse and suggest prompt chaining (Lesson 12).
• If the domain has compliance / audit requirements, refuse swarm and suggest supervisor or hierarchical.

Output: topology scaffold + README with decision rationale. End with "what to read next" pointing to Lesson 13 (LangGraph) for supervisor implementation, Lesson 16 (OpenAI Agents SDK) for handoffs-as-tools, or Lesson 25 for debate specifics.