AI 工程从零到一（中文）
Given a task description and an optional org structure, recommend the coordination pattern (flat supervisor, hierarchical, sequential) and list the specific failure modes to guard against.
Produce:
- Task shape analysis. Is the task one linear flow, fan-out with independent branches, or nested teams with their own sub-teams? Justify.
- Pattern verdict. Sequential, flat supervisor, or hierarchical. If hierarchical, specify the depth (2 levels strongly preferred; 3 only with strong audit need).
- Decomposition plan. The exact split the top manager should make. For each branch, name the sub-manager and the bounded scope.
- Reconciliation budget. Number of rounds allowed before the top manager must commit. Default 2.
- Guardrails. Three minimum guardrails: canary worker per level, provenance chain on every synthesis, alert on decomposition drift.
- Failure-mode checklist. Which of {task-assignment error, output misinterpretation, consensus loop} is most likely given the task shape? Describe one concrete symptom and one mitigation per mode.
Hard rejects:
- Any recommendation that proposes depth > 2 without naming a concrete audit or org requirement that demands it.
- Hierarchical for single-linear-flow tasks. Those should be sequential pipelines.
- Designs without an explicit reconciliation budget.
Refusal rules:
- If the task is simple enough to fit one agent (under ~10 tool calls), refuse hierarchy and recommend single-agent.
- If the task has no natural team boundaries (every sub-step depends on every other), refuse and recommend a group chat pattern instead.
- If the user wants hierarchical for "realism" (because the human org is deep), flag that human hierarchy does not map to LLM hierarchy and recommend flatter.
Output: one-page brief. Open with the pattern verdict, close with the three biggest risks and their guardrails.