feat(agents): migrate createReactAgent to core LangChain project

[christian-bromann]

This patch migrates the createReactAgent from LangGraphJS to LangChain canonical. Here is a list of changes that users can expect from this updated agent:

💥 Breaking Changes

• The createReactAgent doesn't support pre-bound models anymore as it is required to bind tools itself

  const llm = new ChatOpenAI({ model: "gpt-4o" }).bindTools([tool]);
  // ❌ not supported anymore
  const agent = createReactAgent({ llm });

  Instead, users have to pass a along tools to the createReactAgent function.

  const llm = new ChatOpenAI({ model: "gpt-4o" })
  // ✅ supported
  const agent = createReactAgent({ llm });

• The ToolNode doesn't send an Please fix your mistakes. message anymore, instead it will just throw
  directly if the user doesn't handle the error via handleToolError

🚀 New Features

• ToolNode now also takes a callback as argument to catch errors and send custom tool call responses back to the model, e.g.:

  const toolNode = new ToolNode([toolWithError], {
    handleToolErrors: (error, toolCall) => {
      // handle the error and return a custom tool call response
      console.error(error);
      return new ToolMessage({
        content: "handled error",
        tool_call_id: toolCall.id!,
      });
    },
  });
  const agent = createReactAgent({
    llm,
    tools: [toolNode],
  });

• Creating structured outputs is now supported via tool calls or JSON schema (if supported by the model).
  While before the agent would make an additional call to the model to parse the output, now it will just
  return the raw output from the tool call or model directly.

  const llm = new ChatOpenAI({ model: "gpt-4o" })
  const agent = createReactAgent({
    llm,
    // by default, the agent will use tool calls to create structured outputs
    responseFormat: z.object({
      weather: z.string(),
      temperature: z.number(),
    }),
  });
  const result = await agent.invoke({
    messages: [new HumanMessage("What is the weather in Tokyo?")],
  });
  console.log(result.structuredOutput);

  Alternatively, users can also pass a JSON schema to the responseFormat option to create structured outputs.

  const agent = createReactAgent({
    llm,
    responseFormat: z.object({
      weather: z.string(),
      temperature: z.number(),
    }),
  });

  If your model supports JSON schema, you can explicitly request it via the nativeOutput function:

  import { nativeOutput } from "langchain";
  const agent = createReactAgent({
    llm,
    responseFormat: nativeOutput(z.object({
      weather: z.string(),
      temperature: z.number(),
    })),
  });

• When using tool calls to create structured outputs, users now have the ability to catch issues
  with multiple tool calls or invalid parsing by providing a handleError function.

  const agent = createReactAgent({
    llm,
    responseFormat: toolOutput([
      z.object({
        weather: z.string(),
        temperature: z.number(),
      })
    ], {
      handleError: (error) => {
        return `Failed to parse the output: ${error}, please try again.`;
      },
    }),    

createReactAgent API Reference

The createReactAgent function is the main entry point for creating a React agent. It takes in a set of parameters and returns an agent instance. All related types are exported from the langchain package.

• createReactAgent: the main function to create a React agent.
• toolOutput: a function to create a tool output.
• nativeOutput: a function to create a native output.
• AgentState: the state of the agent.
• AgentRuntime: the runtime of the agent.
• CreateAgentToolConfig: the runnable config when accessing runtime information within tools.

import {
  createReactAgent,
  toolOutput,
  nativeOutput,
  type AgentState,
  type AgentRuntime,
  type CreateAgentToolConfig
} from "langchain";

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Preview	Comments	Updated (UTC)
langchainjs-docs	Error		Comment	Aug 27, 2025 10:49pm

1 Skipped Deployment

Project	Deployment	Preview	Comments	Updated (UTC)
langchainjs-api-refs	Ignored			Aug 27, 2025 10:49pm

[christian-bromann]

Blocked on #8793