AI Engineering from Scratch（英文原版）

Given a multi-agent task description, pick the MARL algorithm.

Produce:

1. Task taxonomy. Fully cooperative (shared reward), fully competitive (zero-sum), mixed, general-sum. Number of agents. Homogeneous vs heterogeneous.
2. Observability. Full (every agent sees global state), partial (each sees own observation only), or communication-enabled.
3. Action space. Discrete (Atari-like, SMAC) or continuous (particle world, MuJoCo). Affects algorithm choice.
4. Reward structure. Dense (per-step shaped) vs sparse (terminal only). Dense makes MAPPO practical; sparse needs credit assignment help (QMIX's value decomposition).
5. Algorithm recommendation. Start with MAPPO as baseline per Yu et al. 2022. Switch to:
   • QMIX when cooperative + homogeneous + strong sparse-reward credit assignment needed
   • MADDPG when mixed (cooperative + competitive) + continuous actions
   • Extensions (QTRAN, QPLEX, FACMAC) when monotonicity constraint is too restrictive
6. Training infrastructure. Do you have: enough interaction data, compute budget, reward shaping expertise, stability budget (5-10 seeds per experiment)? If not, recommend prompt-level policies for LLM agents.
7. Deployment contract. CTDE: at deploy time each agent only sees local observation. Write the contract explicitly so runtime code respects it.

Hard rejects:

• Picking a non-MAPPO baseline for a first run. MAPPO is the 2026 baseline; start there.
• Using QMIX for mixed cooperative-competitive tasks. Value decomposition assumes monotone aggregation.
• Recommending MARL training for LLM-agent systems that lack interaction data or reward signal. Prompt-level policies will outperform until the data is there.
• Training without logging per-agent observations and actions. Debugging is impossible.

Refusal rules:

• If the task has fewer than ~1000 episodes of interaction data, recommend prompt-level policies or supervised fine-tuning.
• If the task is non-Markovian (requires memory) but the recommendation does not include recurrent critics, flag the gap.
• If the task is general-sum competitive (multiple equilibria), MARL alone does not pick one; recommend mechanism design or equilibrium selection.

Output: a one-page brief. Start with a one-sentence recommendation ("MAPPO baseline with centralized value function; per-agent discrete actor; CTDE at deploy; 5 seeds per experiment."), then the seven sections above. End with a training-to-deployment pipeline: data collection, training, evaluation, rollout.