Most multi-agent frameworks assume you already know which agent should handle a request — you wire the routing by hand or chain everything through one model. Agent Squad inverts that: the routing decision is the product. A classifier reads the incoming message plus the running conversation, scores every registered agent's description against it, and hands off to whichever fits — so adding a capability means registering an agent, not editing a router.
What Sets It Apart
- Classifier-driven routing weighs both agent profiles and prior turns, so a follow-up like "and cheaper?" still reaches the agent that answered the first question — context survives the handoff.
- Genuinely polyglot: feature-paired Python and TypeScript implementations, not a port lagging behind a primary language.
- Backend-agnostic agents — Amazon Bedrock, Anthropic Claude, OpenAI, Lex, Lambda, or your own class — mix freely behind one classifier.
- The newer SupervisorAgent flips to an "agents-as-tools" model: a lead agent calls specialist teammates in parallel rather than routing to exactly one.
Who It's For
Great fit if you're building an assistant that spans distinct domains (billing, tech support, travel) and want each handled by a focused agent without hand-rolling intent detection. The dual-runtime design suits teams split between Python data stacks and TypeScript services. Look elsewhere if you only need a single ReAct-style agent with tools — the classifier overhead buys you nothing — or if you want a hosted, no-code builder; this is a library you assemble and deploy yourself.
A Note on Provenance
Originally built inside AWS Labs as "Multi-Agent Orchestrator," the project was renamed Agent Squad and has since moved to its original creators' own organization. Treat older awslabs references and the Orchestrator name as the same lineage when reading tutorials or pinned versions.
