Agents

An agent is a model calling tools in a loop until a given task is complete.

Agent = Model + Harness

The job of a harness: get the model the right context at the right time for the given task.

A harness is everything around that loop: the model, its prompt, its tools, and any middleware that shapes its behavior.

create_agent is a highly configurable harness. At its simplest:

import { createAgent } from "langchain";

const agent = createAgent({ model: "openai:gpt-5.4", tools });

Configure the basics directly — model=, tools=, system_prompt=. For more advanced capabilities, extend the harness with middleware.

Core components

Model

Pass a model identifier string ("provider:model") or an initialized model instance. See Models for parameters, provider setup, and dynamic model selection.

import { createAgent } from "langchain";

const agent = createAgent({ model: "openai:gpt-5.4", tools });

Tools

Pass any Python callable, LangChain tool, or tool dict. See Tools for tool definition, context access, and dynamic tool selection.

import { createAgent, tool } from "langchain";
import * as z from "zod";

const search = tool(({ query }) => `Results for: ${query}`, {
  name: "search",
  description: "Search for information",
  schema: z.object({ query: z.string() }),
});

const agent = createAgent({ model: "openai:gpt-5.4", tools: [search] });

System prompt

Shape how the agent approaches tasks. Accepts a string or SystemMessage. For dynamic prompts at runtime, use middleware.

const agent = createAgent({
  model: "openai:gpt-5.4",
  tools,
  systemPrompt: "You are a helpful assistant. Be concise and accurate.",
});

Structured output

Return a validated schema from the agent using response_format=. See Structured output for strategies and examples.

import * as z from "zod";

const Answer = z.object({ summary: z.string(), confidence: z.number() });

const agent = createAgent({ model: "openai:gpt-5.4", tools, responseFormat: Answer });
const result = await agent.invoke({ messages: [{ role: "user", content: "Summarize AI trends" }] });
result.structuredResponse; // { summary: ..., confidence: ... }

Name

Optional identifier used as the node name when embedding this agent as a subgraph in multi-agent systems.

const agent = createAgent({ model: "openai:gpt-5.4", tools, name: "research_assistant" });

To extend the agent’s state schema with custom fields, use state_schema on create_agent or define it via middleware. See Memory for details.

Invocation

Trace each step of this loop, debug tool calls, and evaluate agent outputs with LangSmith. Follow the tracing quickstart to get set up.

You can invoke an agent by passing an update to its State. All agents include a sequence of messages in their state; to invoke the agent, pass a new message along with a thread_id so the agent can persist and resume conversation history:

Google

import { AIMessage } from "@langchain/core/messages";
import { createAgent } from "langchain";
import { MemorySaver } from "@langchain/langgraph";

const agent = createAgent({
  model: "google-genai:gemini-3.5-flash",
  tools: [],
  checkpointer: new MemorySaver(),
});

const config = { configurable: { thread_id: crypto.randomUUID() } };

let result = await agent.invoke(
  {
    messages: [
      { role: "user", content: "What's the weather in San Francisco?" },
    ],
  },
  config,
);

// A follow-up turn on the same conversation: reuse the same thread_id to keep history
result = await agent.invoke(
  { messages: [{ role: "user", content: "What about tomorrow?" }] },
  config,
);

Persisting conversation history with thread_id requires the agent to be configured with a checkpointer. When deployed on LangSmith, a checkpointer is provisioned automatically. Locally, pass one explicitly, for example create_agent(..., checkpointer=InMemorySaver()).

If you also need to pass per-run configuration (such as a user ID, API keys, or feature flags) to tools and middleware, pass it as context alongside the config. Define the shape of that data with contextSchema and access it through runtime.context:

Google

import * as z from "zod";
import { AIMessage } from "@langchain/core/messages";
import { createAgent } from "langchain";
import { MemorySaver } from "@langchain/langgraph";

const contextSchema = z.object({
  user_id: z.string(),
});

const agent = createAgent({
  model: "google-genai:gemini-3.5-flash",
  tools: [],
  contextSchema,
  checkpointer: new MemorySaver(),
});

const result = await agent.invoke(
  {
    messages: [
      { role: "user", content: "What's the weather in San Francisco?" },
    ],
  },
  {
    configurable: { thread_id: crypto.randomUUID() },
    context: { user_id: "user-123" },
  },
);

thread_id scopes the conversation (message history, checkpoints), while context carries per-run data your tools and middleware read at invocation time. Both are commonly passed together. See tool context and Runtime for more.

Streaming

We’ve seen how the agent can be called with invoke to get a final response. If the agent executes multiple steps, this may take a while. To show intermediate progress, we can stream back messages as they occur.

const stream = await agent.stream(
  {
    messages: [{
      role: "user",
      content: "Search for AI news and summarize the findings"
    }],
  },
  { streamMode: "values" }
);

for await (const chunk of stream) {
  // Each chunk contains the full state at that point
  const latestMessage = chunk.messages.at(-1);
  if (latestMessage?.content) {
    console.log(`Agent: ${latestMessage.content}`);
  } else if (latestMessage?.tool_calls) {
    const toolCallNames = latestMessage.tool_calls.map((tc) => tc.name);
    console.log(`Calling tools: ${toolCallNames.join(", ")}`);
  }
}

For more details on streaming, see Streaming.

Configure the harness

create_agent is highly extensible. Middleware is the primitive for customization: each piece handles one concern, hooks into the agent loop at the right moment, and composes freely with any other. Take exactly what your use case needs — skip the rest.

Common patterns are pre-built as first-class middleware. Anything custom is one middleware away.

As agents take on complex work, they need support across a few key areas. The middleware ecosystem covers each:

Execution environment

Tools, filesystem, sandboxes, and code execution

Context management

Summarization, memory, skills, and prompt caching

Planning and delegation

Todo lists and subagents for parallel, isolated work

Fault tolerance

Retries, fallbacks, and call limits

Guardrails

PII detection and content controls

Steering

Human-in-the-loop approval before high-impact actions

Execution environment

Agents are useful when they can take action — not just generate text. The execution environment gives the agent a workspace: tools it can call, a filesystem for reading and writing files across turns, and code execution for running scripts or shell commands.

import { createAgent } from "langchain";
import { FilesystemMiddleware, StateBackend } from "deepagents";

const agent = createAgent({
  model: "anthropic:claude-sonnet-4-6",
  tools: [search],
  middleware: [new FilesystemMiddleware({ backend: new StateBackend() })],
});

See FilesystemMiddleware, Sandboxes, Interpreters.

Context management

Every model call has a fixed context window. As an agent runs — accumulating history, tool results, and intermediate steps — that window fills. Summarization compresses history before overflow hits; memory loads persistent instructions at startup so knowledge carries across sessions; skills surface domain knowledge on demand rather than loading everything upfront.

import { FilesystemMiddleware, MemoryMiddleware, SkillsMiddleware, SummarizationMiddleware, StateBackend } from "deepagents";

const backend = new StateBackend();
const model = "anthropic:claude-sonnet-4-6";

const agent = createAgent({
  model,
  tools: [search],
  middleware: [
    new FilesystemMiddleware({ backend }),
    new SummarizationMiddleware({ model, backend }),
    new MemoryMiddleware({ backend, sources: ["./AGENTS.md"] }),
    new SkillsMiddleware({ backend, sources: ["./skills/"] }),
  ],
});

See SummarizationMiddleware, MemoryMiddleware, SkillsMiddleware, Context engineering.

Planning and delegation

Complex tasks often exceed what one context window can handle. Delegation lets the main agent break work into pieces, hand them to subagents that each run in their own isolated context, and stay focused on coordination rather than execution. Work can run in parallel; the main agent’s context stays clean.

import { createAgent, TodoListMiddleware } from "langchain";
import { FilesystemMiddleware, SubAgentMiddleware, StateBackend } from "deepagents";

const researcher = { name: "researcher", description: "Searches and returns a structured summary.", tools: [search] };

const agent = createAgent({
  model: "anthropic:claude-sonnet-4-6",
  tools: [search],
  middleware: [
    new FilesystemMiddleware({ backend: new StateBackend() }),
    new TodoListMiddleware(),
    new SubAgentMiddleware({ backend: new StateBackend(), subagents: [researcher] }),
  ],
});

See SubAgentMiddleware, Subagents.

Fault tolerance

Agents in production encounter failures that rarely appear in development: rate limits, model timeouts, transient API errors. Fault tolerance middleware handles these at the infrastructure level so your tools and business logic don’t need try/catch around every call.

import { createAgent, ModelRetryMiddleware, ToolRetryMiddleware } from "langchain";

const agent = createAgent({
  model: "anthropic:claude-sonnet-4-6",
  tools: [search],
  middleware: [new ModelRetryMiddleware({ maxRetries: 3 }), new ToolRetryMiddleware({ maxRetries: 2 })],
});

See ModelRetryMiddleware, ToolRetryMiddleware, Prebuilt middleware.

Guardrails

Some policies can’t live in a prompt — they need to be enforced deterministically regardless of what the model does. Guardrails intercept data as it flows through the agent loop, applying compliance rules or content policies before tool results reach the model’s context.

import { createAgent, PIIMiddleware } from "langchain";

const agent = createAgent({
  model: "anthropic:claude-sonnet-4-6",
  tools: [search],
  middleware: [new PIIMiddleware()],
});

See PIIMiddleware, Prebuilt middleware.

Steering

Full autonomy isn’t always appropriate. Steering lets you place humans at specific decision points — before destructive writes, expensive API calls, or anything requiring judgment — without restructuring your agent. The agent pauses and waits; a human approves, edits, or rejects; execution continues.

import { createAgent, HumanInTheLoopMiddleware } from "langchain";

const agent = createAgent({
  model: "anthropic:claude-sonnet-4-6",
  tools: [search],
  middleware: [new HumanInTheLoopMiddleware({ interruptOn: { writeFile: true } })],
});

See HumanInTheLoopMiddleware, Human-in-the-loop.

create_deep_agent pre-assembles this stack for long-running coding and research tasks (filesystem, summarization, subagents, and prompt caching included by default). See Deep Agents for the full pre-built harness.

Middleware resources:

• Middleware overview: how the middleware stack works and when hooks fire
• Prebuilt middleware: full reference with configuration examples
• Custom middleware: write your own hooks for business logic, PII scrubbing, and more

---

Connect these docs to Claude, VSCode, and more via MCP for real-time answers.

Edit this page on GitHub or file an issue.