Agent Squad — Multi-Agent Orchestration for Complex Conversations

Introduction

Agent Squad is an open-source multi-agent orchestration framework that routes user messages to the most appropriate specialized agent. Instead of building one monolithic agent that handles everything, you define multiple focused agents and let the orchestrator intelligently dispatch conversations based on intent classification and context.

What Agent Squad Does

• Routes conversations to specialized agents based on message intent
• Maintains conversation context across agent handoffs
• Supports parallel agent execution for multi-faceted queries
• Integrates with AWS Bedrock, OpenAI, Anthropic, and other providers
• Provides both Python and TypeScript implementations

Architecture Overview

The orchestrator uses a classifier (LLM-based or rule-based) to analyze each incoming message and determine which registered agent should handle it. Agents are defined with names, descriptions, and capabilities. The classifier compares the message against agent descriptions and routes accordingly. Conversation history is maintained per-session, enabling context-aware routing even when topics shift mid-conversation.

Self-Hosting & Configuration

• Install via pip (Python) or npm (TypeScript)
• Configure agents with any LLM provider (Bedrock, OpenAI, Anthropic, local)
• Define agent routing rules via descriptions or custom classifiers
• Deploy as a serverless function, container, or library within your app
• No infrastructure beyond your chosen LLM provider

Key Features

• Intent-based routing across multiple specialized agents
• Conversation context preservation during agent transitions
• Support for both streaming and batch responses
• Provider-agnostic design works with any LLM backend
• Dual implementation in Python and TypeScript

Comparison with Similar Tools

• CrewAI — task-oriented sequential agents; Agent Squad focuses on conversation routing
• AutoGen — multi-agent conversation between agents; Agent Squad routes users to agents
• LangGraph — graph-based workflows; Agent Squad provides simpler dispatcher pattern
• OpenAI Swarm — lightweight handoff protocol; Agent Squad adds classification and context management
• Semantic Kernel — broader AI framework; Agent Squad is focused solely on multi-agent dispatch

FAQ

Q: How does it decide which agent to route to? A: A classifier agent analyzes the message against all registered agent descriptions and picks the best match. You can also provide custom classification logic.

Q: Can one conversation involve multiple agents? A: Yes. As the topic shifts, the orchestrator re-routes to the appropriate agent while preserving context.

Q: Does it work with local LLMs? A: Yes. Any OpenAI-compatible endpoint works, including Ollama and vLLM.

Q: What happens if no agent matches? A: You can configure a default fallback agent or return a clarification request to the user.

Sources

• https://github.com/2FastLabs/agent-squad