Description
Agent Squad (formerly multi-agent-orchestrator, and previously hosted under AWS Labs) is an open-source framework for building systems made of several specialized AI agents instead of one general-purpose chatbot. A classifier reads the user's query and the conversation history and picks the most suitable agent to handle that turn; an orchestrator then keeps context flowing between agents so a multi-turn conversation stays coherent even as it's routed to different agents.
Read More
The project ships as three runtimes that the maintainers describe as feature-parity: a TypeScript package (npm install agent-squad) and a Python package (pip install agent-squad, with extras for aws, anthropic, openai or all), both of which run anywhere Node or Python run, including AWS Lambda or a container. A newer Swift package brings the same orchestration model to iOS and macOS, running entirely on-device with native and MCP (Model Context Protocol) tool calling, realtime voice, tracing and local-first chat storage.
Out of the box, Agent Squad includes pre-built agent wrappers for Amazon Bedrock, Anthropic, OpenAI, Amazon Lex and AWS Lambda, plus built-in classifiers, storage backends and retrievers, with defined extension points for custom agents, classifiers or storage. Responses can stream or return in full. A newer addition, JevClassifier, replaces the default LLM-based classifier with a decision model (Jev, from TypeSafe AI) that returns the winning agent plus a calibrated confidence score in under a second, billed only on input tokens, and supports an explicit "unknown" outcome for out-of-scope requests instead of forcing a match.
The repository was transferred from awslabs/agent-squad to 2FastLabs/agent-squad and was originally released as multi-agent-orchestrator; the change of GitHub org signals a change of maintainer from AWS Labs to an independent team, though the framework and its APIs continue from the same codebase. It's licensed Apache 2.0 and has around 7,700 GitHub stars at the time of writing.
Agent Squad suits developers building chatbots or assistants that need to hand off between several distinct skills or knowledge domains (for example a tech-support agent versus a billing agent) rather than relying on one prompt to do everything, and teams that want the same orchestration logic available in a backend language (Python or TypeScript) and natively on Apple devices.
- CategoryAI Agent Builders
- Pricing ModelOpen source
- DeploymentUniversal (Cloud, Local, AWS Lambda)
- PlatformsWeb, iOS, macOS
Agent Squad's Core Features
Classifier-based routing of each query to the most suitable specialized agent
Conversation context maintained across agents and sessions
Streaming and non-streaming agent responses
Pre-built agents for Bedrock, Anthropic, OpenAI, Amazon Lex and AWS Lambda
Extensible storage, classifier and retriever interfaces for custom setups
Feature-parity Python and TypeScript runtimes, usable in Lambda, containers or locally
On-device Swift runtime for iOS and macOS with native and MCP tool calling
Realtime voice and local-first chat storage in the Swift runtime
JevClassifier: sub-second, calibrated-confidence routing via a decision model instead of an LLM call
Apache 2.0 license, maintained at 2FastLabs/agent-squad (formerly AWS Labs' multi-agent-orchestrator)
Getting Started with Agent Squad
Install the package: npm install agent-squad (TypeScript) or pip install "agent-squad[aws]" (Python).
Create an orchestrator: Instantiate AgentSquad() in your app.
Add agents: Register pre-built agents (e.g. BedrockLLMAgent) with a name and description of their specialty.
Route requests: Call routeRequest()/route_request() with the user's message, user ID and session ID.
Handle the response: Read response.output (or stream chunks) and response.metadata.agentName to see which agent answered.
Swap in JevClassifier: Pass classifier: new JevClassifier() for faster, decision-model-based routing.
Agent Squad's Use Cases
- Multi-skill chatbots
- On-device iOS/macOS assistants
- AWS-native agent backends
- Fast intent routing
