build-agent

To build an LLM agent with tool calling, you create an LlmAgent configuration using a LangChain chat model, register tool functions, set instructions, and run it through a session service to stream events reliably across async sessions.

To enforce structured reasoning in an LLM agent, you use a ReActAgent configuration which enforces typed ReAct steps at each iteration via structured output, ensuring reliable and typed reasoning throughout the execution loop.

You manage per-session state when orchestrating LLM agents by running the configured agent through a Runner-managed session service, which handles state persistence and streams events back during the execution loop.

Yes, you can use LangChain tools with a ReActAgent by registering tool functions like web search during agent configuration, allowing the agent to execute tool calls in an iterative loop while enforcing typed ReAct steps.

Yes, async streaming supports structured outputs by utilizing BaseAgent-compatible astream behavior returning AsyncIterator[Event], allowing you to parameterize models, tools, and instructions while getting typed reasoning results.

Configuring max iterations in an LLM agent loop limits the number of tool execution cycles the agent performs, requiring careful parameterization of models and instructions to ensure the agent completes its reasoning within the set boundary.