Swarm Agent
A collection of peer agents with equal authority that coordinate through direct handoffs and shared context. No central controller — agents decide when to hand off to another agent based on the current task. Coordination is emergent, not enforced.
Structure
Any agent can receive the initial task. Agents hand off to peers when a task falls outside their scope. The active agent has full control until it explicitly transfers to another. Multiple agents may contribute to the final result.
Memory / State / Context
- Memory
- State
- Context
- Typically short-lived per-agent memory
- Heavy reliance on summarization during handoffs
- Shared message pool or blackboard for coordination
- No single agent owns the complete memory
- Shared context variable space across agents
- State transferred explicitly during handoffs
- No central state tracker — state is distributed
- Each agent can read and modify shared variables
- Shared problem statement and collaboration rules
- Active agent receives full conversation history on handoff
- Context grows as agents add to the shared history
- Risk of context bloat without summarization
Key Characteristics
- High flexibility — agents self-organize around the task
- No single point of failure — any agent can take over
- Emergent behavior — solutions arise from agent interactions
- Low determinism — same input may produce different execution paths
- Hard to debug — no clear execution trace or authority chain
When to Use
- Tasks require dynamic routing between multiple specialized capabilities
- No single agent can predict what expertise is needed upfront
- You want lightweight coordination without a central orchestrator
- The problem benefits from peer-to-peer handoffs (like customer service triage)
- You're building conversational systems where context transfers between specialists