AI 原生开放教材

外观

AI 工程从零到一（中文）源码研读·中文原文 ↗原件 ↓仓库 ↗

AI 工程从零到一（中文） ​

Given a user query that benefits from parallel subagent research, produce a supervisor-pattern design ready to wire into any framework (LangGraph, OpenAI Agents SDK, CrewAI Hierarchical).

Produce:

  1. Complexity estimate. Is this query simple (1 agent, 3-10 tool calls), medium (2-4 workers), or complex (5+ workers)? Justify in one sentence using Anthropic's scale-effort heuristic.
  2. Lead system prompt. Must include: (a) decomposition instructions, (b) synthesis instructions, (c) explicit rule that the lead never reads raw source content, only worker summaries.
  3. Worker system prompts. One per role, each naming its narrow scope and the output format the lead expects.
  4. Sub-question decomposition rules. How does the lead split the query? Broad-first-then-narrow, or direct decomposition? What disqualifies a sub-question (overlap with another, too broad)?
  5. Synthesis template. Explicit conflict-handling rule: if two workers return contradictory facts, the synthesis must surface the disagreement rather than silently picking one.
  6. Model pairing. Which model for the lead (reasoning tier), which for workers (faster/cheaper tier). Explain the tradeoff.
  7. Observability requirements. Minimum trace points: plan, each worker start/end, synthesis input, synthesis output.

Hard rejects:

  • Any design where the lead does tool-use itself. Lead only plans and synthesizes.
  • Worker prompts that permit scope drift (e.g., "research anything related to X" without a bound).
  • Synthesis templates that hide conflicts.

Refusal rules:

  • If the query is simple (estimated under 10 tool calls total), refuse the design and recommend single-agent instead. Cite the Anthropic 15× token cost finding.
  • If the query is sequential (step 2 needs step 1's output), refuse and recommend a pipeline/chain pattern instead.
  • If the user is optimizing for determinism and audit, refuse supervisor and recommend a LangGraph static graph.

Output: one-page design brief. Start with the complexity estimate and a pattern-fit verdict ("supervisor fits"). Close with a rainbow-deployment reminder if the system will run continuously.