AI Engineering from Scratch（英文原版）
Given a user-facing task (often triage or skill-based routing), produce a handoff topology ready to map onto OpenAI Swarm or the OpenAI Agents SDK.
Produce:
- Agent roster. Each agent: name, one-sentence purpose, tools, and which other agents it can hand off to.
- Handoff functions. The tool signatures per agent. Each handoff function returns a target Agent.
- Context transfer policy. On each handoff edge: full history, last N messages, or summarized snapshot. Justify.
- Guardrails. Input validation per agent (what prompts are allowed to trigger handoffs to sensitive specialists), authentication on handoff where needed.
- Loop detection. Rule to detect ping-pong (e.g., "A handed off to B; B handed off back to A" occurring more than once in a row).
- Fallback behavior. If a handoff target is missing (removed agent, auth failure), which agent handles the session.
- Session / memory plan. Whether to use Agents SDK sessions, caller-managed memory, or no memory at all.
Hard rejects:
- Any handoff design without loop detection.
- Handoff functions that pass full history to specialists with different tool permissions (security risk).
- Designs that assume Swarm's stateless behavior but then require multi-turn memory — use Agents SDK sessions instead.
Refusal rules:
- If the task needs parallel execution, refuse Swarm and recommend supervisor (Lesson 05) instead.
- If the task needs deterministic audit/replay, refuse and recommend LangGraph static graph.
- If the task is a simple DAG of stages (research → code → review), recommend CrewAI Sequential instead.
Output: a one-page handoff brief. Close with a security note on how prompt injection could trigger unwanted handoffs and what guardrails block it.