Multi-Agent Swarms

Multi-Agent Swarms

Swarms are groups of agents working in parallel on decomposed tasks. AgentCTX’s orchestrator manages spawn, coordination, and result aggregation.

Concepts

Orchestrator
├── Agent A (frontend)    ← Phase 1
├── Agent B (backend)     ← Phase 1
├── Agent C (tests)       ← Phase 2 (depends on A + B)
└── Agent D (docs)        ← Phase 2 (depends on A + B)

Key properties:

• Phased execution — tasks execute in dependency order
• Parallel within phases — agents in the same phase run concurrently
• Coherence detection — prevents agents from duplicating work
• Shared memory — all agents in a swarm share team-level memory

Launching a Swarm

Via CLI

actx swarm manifest.yaml

Manifest Format

name: feature-implementation
phases:
  - name: implementation
    agents:
      - role: frontend
        prompt: "Implement the login page component"
        model: claude-sonnet
      - role: backend
        prompt: "Implement the auth API endpoint"
        model: claude-sonnet

  - name: validation
    depends_on: [implementation]
    agents:
      - role: tester
        prompt: "Write integration tests for login flow"
        model: gemini-flash
      - role: reviewer
        prompt: "Review both frontend and backend changes"
        model: claude-opus

Swarm Lifecycle

1. Decompose — break the task into parallelizable work items
2. Dispatch — spawn agents for each work item
3. Monitor — track progress via actx top or actx dashboard
4. Coordinate — agents share state through team memory
5. Aggregate — collect results and resolve conflicts
6. Report — generate summary of all agent work

Monitoring

# Real-time process tree
actx top

# Dashboard with metrics
actx dashboard

The dashboard shows:

• Agent status (running, waiting, complete, failed)
• Phase progress (which phases are active)
• Token budget consumption per agent
• Coherence alerts (duplicate work detected)

Memory Coordination

Agents in a swarm communicate through team-level memory:

// Agent A stores a decision
+m "auth-approach" #team "Using PASETO tokens"

// Agent B reads team decisions before starting
?m #team @1h ^10

// Agent C confirms the approach
~m auth-approach #team +confirm

Error Handling

Scenario	Behavior
Agent fails	Retry with same prompt, then escalate
Budget exceeded	Pause agent, notify orchestrator
Coherence conflict	Merge or ask human to resolve
Phase dependency fails	Block downstream phases

See Also

• Multi-Agent Orchestration — orchestrator architecture
• Agent-to-Agent Communication — messaging patterns
• actx top — process monitor