// the find
2FastLabs/agent-squad
Flexible and powerful framework for managing multiple AI agents and handling complex conversations
Agent Squad is a multi-agent orchestration framework that routes user queries to the most appropriate specialized agent using an LLM-based classifier. Originally an AWS Labs project (multi-agent-orchestrator), it was recently moved to 2FastLabs and supports Python and TypeScript, with built-in integrations for AWS Bedrock, Anthropic, OpenAI, and Lex. It's aimed at developers building chatbots or AI pipelines that need to dispatch across multiple specialized models or services.
The classifier-based routing is the real value here — instead of manually wiring conditionals, the framework uses an LLM to pick the right agent based on query semantics and conversation history, which handles ambiguous inputs better than keyword matching. The modular install options (pip install agent-squad[anthropic] vs [aws]) are thoughtfully designed and avoid forcing Boto3 on everyone who just wants Anthropic. The SupervisorAgent implementing agents-as-tools for parallel subtask delegation is genuinely useful for workflows that need fan-out and aggregation. Both Python and TypeScript are first-class citizens with parity, not an afterthought port.
The AWS/Bedrock bias runs deep — the demo app, the CDK examples, the primary storage options — so if you're not on AWS, you're fighting the grain of the design rather than using it. The conversation storage story is weak for production: in-memory by default, DynamoDB as the main durable option, SQL support exists but is thin and underdocumented. The repo-rename from awslabs means badges and star counts still point at the old org, which signals uncertain maintenance trajectory. The classifier makes an LLM call on every user message, which adds latency and cost even for simple single-agent setups where static routing would be faster and cheaper.