Full-Code · Orchestration Frameworksactive

Agent Squad

also known as Multi-Agent Orchestrator, awslabs/agent-squad

Type: full-code · Vendor: 2FastLabs (originally AWS Labs) · Language: Python / TypeScript / Swift · License: Apache-2.0 · Status: active · Status in practice: mature

Links: homepage docs repo

Lightweight open-source multi-agent orchestration framework: a classifier routes each user query to the most suitable specialized agent while conversation context is maintained across agents; a SupervisorAgent adds team coordination via an agent-as-tools architecture.

Description. Agent Squad (formerly multi-agent-orchestrator, previously hosted at awslabs/agent-squad, now maintained at 2fastlabs/agent-squad) orchestrates multiple specialized agents behind a single conversational surface. Every user turn is analyzed by a Classifier that uses agent descriptions plus conversation history to select the best agent; the Orchestrator dispatches the turn, saves the exchange, and returns the response. A built-in SupervisorAgent implements an agent-as-tools architecture in which team members are exposed to a supervisor as invocable tools, enabling parallel processing and contextual communication. Python and TypeScript runtimes maintain feature parity and run anywhere from AWS Lambda to a laptop; a Swift runtime brings the same orchestration model — classifier routing, native and MCP tools, realtime voice, tracing — fully on device for Apple platforms.

Agent loop shape. User input -> Classifier (agent descriptions + history) -> selected agent processes the turn -> Orchestrator saves the exchange and returns the response; optionally a SupervisorAgent fans work out to team agents exposed as tools and aggregates their results.

Primary use cases

routing one conversational surface across many specialized agents
supervisor-coordinated agent teams
serverless and on-device multi-agent deployment

flowchart TD U[User input] --> C[Classifier] C -->|best agent for turn| A1[Specialized agent] C --> A2[Specialized agent] A1 --> O[Orchestrator saves exchange] A2 --> O S[SupervisorAgent] -->|agents as tools| T1[Team agent] S --> T2[Team agent] T1 --> S T2 --> S S --> O O --> R[Response]

Key concepts

Classifier (docs) — Analyzes each user turn against agent descriptions and conversation history to pick the best agent.
SupervisorAgent (docs) — Orchestration component exposing team members to a supervisor as invocable tools for parallel, context-aware coordination.
Swift on-device runtime (docs) — Same orchestration model — classifier routing, native and MCP tools, realtime voice, tracing — running entirely on device.