Agent Squad

An open-source AI agent framework routes conversations across specialized agents, with Python and TypeScript runtimes and on-device Swift orchestration.

Agent Squad is a framework for developers building assistants that divide work among specialized AI agents. It handles agent selection and conversation context so users can switch topics without starting a separate chat.

Conversation routing: Select agents using the request and chat history, preserve context across handoffs, and return streamed or complete responses.
Agent teams: SupervisorAgent delegates subtasks to several agents in parallel and combines their work into one response.
Data-grounded replies: GroundedAgent separates tool-based data gathering from reply writing, restricting the reply writer to curated tool results for questions about prices, balances or availability.
Custom integrations: Use built-in agents for Amazon Bedrock, Anthropic, OpenAI, Amazon Lex and AWS Lambda, or supply your own agents, classifiers, storage and retrievers.
Apple app capabilities: The Swift runtime supports native and MCP tools, spoken conversations with interruption, and chat history stored on the device.

Python and TypeScript support the same features and can run on your laptop, in containers or on AWS Lambda. Swift runs orchestration on iPhone, iPad and Mac, with iOS 16+ or macOS 14+. Local orchestration doesn't make cloud-backed agents offline: OpenAI, Anthropic and Bedrock calls send requests to those services.

The optional JevClassifier for Python and TypeScript uses TypeSafe AI's external Jev service for routing, returning confidence scores and allowing requests to remain unmatched. Agent Squad is open source under the Apache 2.0 license.