Multi-Agent Overview

LeafMesh is designed for multi-agent orchestration. Multiple agents work together through the managed mesh, with the control plane handling routing, validation, and coordination.

Core Concepts

Agents are the data plane: They execute tasks and produce structured outputs
The mesh is the control plane: It routes traffic, enforces policies, and maintains state
Agents never communicate directly: Every interaction crosses the control plane boundary

Architecture: MANAGED_MESH

All LeafMesh deployments use the MANAGED_MESH architecture:

                    ┌─────────────────────────────────────┐
                    │           CONTROL PLANE              │
                    │                                      │
                    │  Routing       Condition evaluation  │
                    │  Coordination  Self-healing          │
                    │  Events        Scheduling            │
                    │                                      │
                    └────┬──────┬──────┬──────┬───────────┘
                         │      │      │      │
                    ┌────┴──┐ ┌─┴───┐ ┌┴────┐ ┌┴──────┐
                    │Agent A│ │Agent│ │Agent│ │Human  │
                    │ (LLM) │ │  B  │ │  C  │ │Agent D│
                    └───────┘ └─────┘ └─────┘ └───────┘
                              DATA PLANE

Agent Communication Flow

agents:
  triage:
    name: "triage"
    yields:
      category: "string"
      urgency: "number"
    can_call:
      - agent: "specialist"
        condition: "urgency >= 7"
      - agent: "general"
        condition: "urgency < 7"

  specialist:
    name: "specialist"
    yields:
      analysis: "string"
    can_call:
      - agent: "human_review"
        condition: "true"

  general:
    name: "general"
    yields:
      response: "string"

When triage produces urgency: 8:

The condition urgency >= 7 matches
The platform routes the call to specialist
Specialist's yields are passed to human_review
The coordinator observes every event and intervenes only when needed
Session state and yields are persisted automatically

Mixed Agent Types

A single system can combine all four agent types:

Type	Purpose	Example
LLM	AI-powered processing	Triage, analysis, content generation
Programmatic	Deterministic logic	Data validation, calculations
Human	Manual review	Approvals, quality checks
Scheduled	Time-based activation	Monitoring, reports, maintenance

Coordination Components

The control plane provides automatic coordination:

Coordinator: Observes every event and intervenes deterministically -- retry, halt, or escalate -- based on coordination rules you configure.
Self-healing: Detects failures and executes graduated recovery actions.
Event delivery: Real-time event propagation between components, with durable history for replay and audit.

Next Steps

Mesh Architecture — How the mesh works
Agent Communication — Routing and data flow
Coordination Patterns — Automatic oversight and intervention
Event Listeners — Trigger agents from Kafka, SQS, MQTT, Redis Streams, IMAP
Scaling — Scaling multi-agent systems

LeafMesh — Multi-agent orchestration through managed mesh