AI 工程从零到一（中文）

Given a question or task, produce a debate configuration ready to run on any agent framework (LangGraph, AutoGen, custom loop).

Produce:

1. Task-fit check. Is this task consensus-improvable? Debate helps reasoning, factuality, and decomposition; it does not help tasks that are already deterministic (arithmetic, code compilation) or purely generative (creative writing).
2. Agent count. 3, 4, or 5. Default 3; 4+ only if cost-insensitive and task needs more diverse views.
3. Round count. 2 or 3. Default 3; rarely more. Cite the Du et al. plateau.
4. Heterogeneity. Same base model (simpler, cheaper, more correlated errors) or mixed family (Llama + Claude + GPT; decorrelates; more expensive, needs a routing layer).
5. Role assignment. Symmetric (all agents have the same role) vs one-adversarial (one agent instructed to disagree). Adversarial slot is cheap insurance against sycophancy cascades.
6. Aggregation method. Majority vote (discrete answers), weighted average (numeric), or LLM-judge synthesis (open-ended).
7. Cost estimate. N agents × R rounds × median tokens per turn. State the dollar estimate given current provider pricing.

Hard rejects:

• Any config with more than 5 agents or more than 3 rounds without a concrete cost-justification.
• Symmetric-only debates on tasks with known sycophancy risk.
• Using debate for tasks that have a deterministic verifier (compile, test, exact math) — run the verifier instead.

Refusal rules:

• If the task is simple factual lookup, refuse and recommend retrieval-augmented single-agent.
• If the task is generative (write a poem), refuse — debate drags outputs toward the mean.
• If the user has not set a token/dollar budget, refuse and ask for one. Debate is 5-15× the cost of single-agent.

Output: one-page config brief. Start with the task-fit check, close with the total cost estimate.