agent-squad-typescript

Create an AgentSquad orchestrator, register agents with addAgent, and call routeRequest with the user input, userId, and sessionId. A classifier (BedrockClassifier by default, or Anthropic/OpenAI variants) selects the most suitable agent per turn.

Use the async factory MCPToolProvider.create with stdio or sse server configs, then pass the provider as toolConfig.tool to a BedrockLLMAgent. Install the optional peer dependency @modelcontextprotocol/sdk and await create before the agent handles requests.

Yes, agents configured with streaming: true return a Node.js Transform stream in response.output. Check response.streaming and iterate with for await; the transform accumulates the full response internally for storage.

The agent id is derived from its name at construction by stripping non-alphanumerics, lowercasing, and replacing spaces with hyphens. Renaming changes the id used as the storage key, so keep agent names stable across deployments.

GroundedAgent is a two-LLM anti-hallucination pattern where a gatherer agent calls tools and an isolated presenter agent answers only from curated results. Use it when responses must not drift from live data such as prices or balances.

Yes, but you must pass an explicit classifier since AgentSquad defaults to BedrockClassifier, which fails without AWS credentials. Use AnthropicClassifier or OpenAIClassifier with AnthropicAgent or OpenAIAgent instead.