Skip to content

agent-routing.md

Latest commit

 

History

History
88 lines (68 loc) · 2.79 KB

agent-routing.md

File metadata and controls

88 lines (68 loc) · 2.79 KB
Raw

Agent Message Routing Implementation

What We Built

We've successfully implemented actual agent message routing in Pattern! Messages are now sent to real Letta agents based on the routing rules.

Key Components

  1. MultiAgentSystem::send_message_to_agent()

    • Takes user ID, optional agent ID, and message
    • Creates or retrieves the appropriate Letta agent for that user
    • Sends the message and returns the agent's response
    • Each user gets their own instance of each agent type

  2. Agent Creation

    • Agents are created with format: {agent_id}_u{user_id} (e.g., entropy_u123)
    • Each agent is initialized with:
      • Its specific system prompt from AgentConfig
      • Shared memory blocks as context
      • User-specific persona

  3. Message Flow

    Discord/MCP → parse_agent_routing() → send_message_to_agent() → Letta API → Agent Response

Usage Examples

Discord DM:

User: @entropy help me break down writing a report
Entropy: I'll help you break down that report into manageable chunks...

User: flux: when should I schedule this?
Flux: Looking at typical ADHD time patterns, a "2 hour" report will likely need 4-6 hours...

User: hey pattern
Pattern: Hello! I'm here to help coordinate your cognitive support. What do you need today?

MCP Tools:

{
  "tool": "chat_with_agent",
  "arguments": {
    "user_id": 123,
    "agent_id": "momentum",
    "message": "How's my focus looking today?"
  }
}

Agent Routing Rules

  1. Discord formats:

    • @agent_name message - Route to specific agent
    • agent_name: message - Route to specific agent
    • /agent agent_name message - Route to specific agent
    • Plain message - Routes to default agent (usually Pattern)

  2. Agent names are dynamic - based on what's configured in MultiAgentSystemBuilder

Next Steps: Reimplement in pattern

  1. Migrate to Groups API

    • Replace individual routing with group-based coordination
    • Create flexible group configurations
    • Leverage native multi-agent conversation support
    • See Memory and Groups Architecture

  2. Enhanced Memory System

    • Implement three-tier memory hierarchy
    • Add passive knowledge sharing via sources
    • Build cost-optimized sleeptime processing

  3. Specialized Agent Behaviors

    • Entropy: Task breakdown with ADHD awareness
    • Flux: Time translation and reality checks
    • Archive: Memory consolidation and pattern detection
    • Momentum: Energy state tracking
    • Anchor: Gentle habit support

  4. Flexible Coordination Patterns

    • Experiment with different group managers
    • Create context-specific agent groups
    • Build overlapping groups for different needs

The foundation is now in place - next step is leveraging Letta's native multi-agent support!