j4flmao Multi-Agent Systems (MAS)
1. Skill Context
Focus: Designing systems where multiple autonomous AI agents collaborate, argue, or sequentialize tasks to solve complex problems that a single LLM prompt cannot handle. Triggers: multi-agent, swarm, langgraph, autogen, orchestration, state-machine.
2. The Multi-Agent Philosophy
A single LLM acting as a "God Agent" with 50 tools will eventually fail. The context window becomes polluted, and the model forgets its original objective.
Multi-Agent Systems (MAS) solve this by specializing. You create a CoderAgent, a ReviewerAgent, and a QA_Agent, each with distinct system prompts and narrowly scoped tools.
3. Orchestration Architectures
A. State Graphs (Deterministic Routing)
Frameworks: LangGraph The system is explicitly modeled as a Directed Cyclic Graph (DCG).
- Nodes: Represent the Agents (or python functions).
- Edges: Represent the conditional logic routing the flow from one agent to another.
- Pros: Highly predictable, easy to debug, guarantees the workflow will eventually terminate or follow business rules.
- Cons: Rigid. If the user asks for something outside the predefined graph flow, the system cannot adapt dynamically.
B. Swarm Intelligence (Dynamic Orchestration)
Frameworks: AutoGen, OpenAI Swarm Agents act autonomously without a hardcoded graph.
- A user submits a complex request.
- The Manager Agent broadcasts the request to a pool of specialized agents.
- Agents dynamically volunteer to handle parts of the task, pass messages directly to each other, and decide organically when the task is complete.
- Pros: Incredibly flexible. Can solve novel problems the developer never anticipated.
- Cons: Prone to infinite conversational loops ("No, you do it", "No, you do it") and hallucinations. Extremely difficult to debug in an enterprise production environment.
4. References
references/communication-protocols.md— How agents share data (Blackboard vs. Actor Model).references/human-in-the-loop.md— Pausing agent execution for human approval.