Agent Squad

Agent Squad is a router. The AgentSquad orchestrator receives a message with a user id and session id, asks a classifier which registered agent fits best (using each agent's description and the recent conversation), runs that one agent, stores the exchange, and returns the response with metadata naming the agent. If no agent matches, a configurable default agent can answer instead.

Agents do not talk to each other in the basic flow; each turn goes to exactly one of them. Two built-in agent types add coordination inside a turn: SupervisorAgent calls team members as tools, in parallel, and composes one reply, and ChainAgent passes output through a sequence of agents. A supervisor can be registered with the classifier like any agent to build hierarchies.

State is conversation history keyed by user, session and agent, held in a ChatStorage backend (in-memory by default; DynamoDB, SQL or a summarizing wrapper for longer histories). MAX_MESSAGE_PAIRS_PER_AGENT caps how much history each agent keeps.

No built-in approval or pause API is documented. The docs show human involvement as an application pattern: the e-commerce support simulator example routes complex or sensitive queries to human support staff, notifies the customer, and gives staff a separate interface to answer. Any gating of tool calls would have to be written into custom agents or tools.

pip install "agent-squad[aws]"

import asyncio

from agent_squad.agents import BedrockLLMAgent, BedrockLLMAgentOptions
from agent_squad.orchestrator import AgentSquad

# Default classifier and these agents use Amazon Bedrock: configure AWS credentials first.
squad = AgentSquad()
squad.add_agent(BedrockLLMAgent(BedrockLLMAgentOptions(
    name="Billing Agent", description="Answers invoice, refund and payment questions.")))
squad.add_agent(BedrockLLMAgent(BedrockLLMAgentOptions(
    name="Tech Agent", description="Troubleshoots login, app and connectivity problems.")))


async def main():
    resp = await squad.route_request("I was charged twice this month", "user-1", "session-1")
    print(resp.metadata.agent_name, "->", resp.output.content)


asyncio.run(main())

• Python 3.11+ is required for the Python package.
• Pick extras to match providers: agent-squad[aws], [anthropic], [openai] or [all].
• Without a classifier argument, AgentSquad() uses the Bedrock classifier, so AWS credentials and Bedrock model access are needed; the quickstart checks this with aws sts get-caller-identity.
• The project moved from awslabs to 2FastLabs; the README asks users to update bookmarks, clone URLs and dependencies.
• JevClassifier calls a third-party service (TypeSafe) and needs TYPESAFE_API_KEY.
• Pass stream_response=True to route_request and consume response.output as a stream when an agent streams.

The README says the project, previously hosted at awslabs/agent-squad and earlier named multi-agent-orchestrator, is now maintained at 2fastlabs/agent-squad. Package names on PyPI and npm are agent-squad.

Yes. MCPToolProvider connects agents to MCP servers over stdio, Streamable HTTP or SSE in Python and TypeScript, and the Swift runtime supports MCP tools too. A2A and AG-UI support were not found.

Yes, it is Apache-2.0 licensed with no paid tier; the project accepts GitHub sponsorships. Model calls (Bedrock, OpenAI, Anthropic) and the optional Jev classifier service are billed by those providers.

A classifier reads the agents' descriptions and the conversation history and picks one agent per turn. If none fits, a default agent can answer if you enable that option.

It does not drive Claude Code. Each runtime ships a SKILL.md guide for AI coding assistants, and the Swift docs explain installing it as a Claude Code skill so the assistant writes correct Agent Squad code.