Multi-agent systems

What it is

Coordinate multiple agents with separate state, goals, tools, or roles while making communication, contention, and correlated failure visible.

This page separates the concept, implementation path, interactive experiment, evaluation criteria, failure modes, and sources so the topic can be studied without relying on a summary tile.

Core concepts

• Each agent has local observations and internal state.
• Coordination may be cooperative, competitive, or mixed.
• Communication protocols determine what information is shared and when.
• Shared resources create contention and scheduling problems.
• More agents increase both parallelism and the number of possible interactions.

Architecture / workflow

1. Define roles, authority, shared resources, and communication rules.
2. Start with deterministic agents before adding learned policies.
3. Log every message and resource transfer.
4. Test cooperation, conflicting goals, unavailable agents, and adversarial messages.
5. Measure whether multiple agents outperform a simpler single-agent baseline.

Technical foundations

• Multi-agent environments can be modeled as stochastic games or decentralized partially observable MDPs.
• Credit assignment asks which agent/action caused a team outcome.
• Consensus and coordination protocols trade communication cost against consistency.
• Centralized training with decentralized execution is one common design pattern.

Evaluation checklist

• Team task completion and individual contribution.
• Communication volume and latency.
• Resource contention and deadlock rate.
• Robustness when one agent fails or sends bad information.
• Comparison with a single-agent or centralized baseline.

Failure modes and limits

• Agents can amplify one another’s mistaken assumptions.
• Coordination overhead can exceed parallelism benefits.
• Shared memory can become a source of stale or conflicting state.
• Unclear authority can lead multiple agents to perform the same consequential action.

Practical build path

1. Simulate two deterministic agents sharing a resource.
2. Add logged messages and explicit ownership transfer.
3. Introduce conflicting goals and failure injection.
4. Add learned or language-model policies only after the coordination protocol is testable.

Primary / authoritative references

• PettingZoo documentation ↗
• OpenSpiel documentation ↗

Search the Learn library for related lessons →