LangChain DeepAgents

The SDK provides deepagents BaseSandbox implementations for command execution and file transfer. Use ContreeSandbox with a sync executor, or ContreeAsyncSandbox with an async executor. A session implements the matching contract.

Async snippets with top-level await run inside an async function or a notebook that supports it. For a standalone script, use the asyncio.run(main()) structure from Your first sandbox.

The examples use the default sandbox from create_sandbox_session. Commands and uploads use the same VM. Downloads automatically save its current filesystem before reading; this stops the VM. The next command starts from the saved image. The factory handles connection setup and cleanup.

Install and test the backend

This integration requires Python 3.11 or newer. Install the development API from the repository root, as in Your first sandbox:

pip install -e ".[langchain,async]"

The extra pins deepagents 0.7.10, the version tested with this SDK. Its async file tools delegate to aexecute and aupload_files. Version 0.6.8 instead calls sync methods from those tools and is incompatible with ContreeAsyncSandbox. The sandbox image needs a POSIX shell, Python 3, and the file utilities used by deepagents, including grep for search. Replace the example tag:python:3.12-alpine with an image available on your endpoint. Configure a connection as described in Connection sources. This example exercises the backend directly, so no model credentials are needed.

Sync
Async

from contree_sdk import create_sandbox_session
from contree_sdk.langchain import ContreeSandbox


with create_sandbox_session(image="tag:python:3.12-alpine") as session:
    sandbox = ContreeSandbox(session)
    result = sandbox.execute("echo hello")
    print(result.output, end="")
    print("exit:", result.exit_code)

from contree_sdk import create_sandbox_session
from contree_sdk.langchain import ContreeAsyncSandbox


async with create_sandbox_session(image="tag:python:3.12-alpine") as session:
    sandbox = ContreeAsyncSandbox(session)
    result = await sandbox.aexecute("echo hello")
    print(result.output, end="")
    print("exit:", result.exit_code)

The adapter runs commands with disposable=False. Its next tool call sees the filesystem changes. It does not close the session, client, or store. Use SQLite when the agent needs to resume after a Python restart.

Connect the backend to an agent

Pass the sandbox as backend= to deepagents.create_deep_agent. Supply the model configured by your application. Keep the client open for the entire agent invocation. The following function accepts that model and uses the connection configured in Connection sources:

Sync
Async

from deepagents import create_deep_agent

from contree_sdk import create_sandbox_session
from contree_sdk.langchain import ContreeSandbox


def ask_agent(model):
    with create_sandbox_session(image="tag:python:3.12-alpine") as session:
        agent = create_deep_agent(model=model, backend=ContreeSandbox(session))
        request = {"messages": [{"role": "user", "content": "Create a calculator script and run it."}]}
        return agent.invoke(request)

from deepagents import create_deep_agent

from contree_sdk import create_sandbox_session
from contree_sdk.langchain import ContreeAsyncSandbox


async def ask_agent(model):
    async with create_sandbox_session(image="tag:python:3.12-alpine") as session:
        agent = create_deep_agent(model=model, backend=ContreeAsyncSandbox(session))
        request = {"messages": [{"role": "user", "content": "Create a calculator script and run it."}]}
        return await agent.ainvoke(request)

Model-provider installation, credentials, and model selection belong to the agent application. The SDK does not choose a provider or supply a model key.

Advanced: file transfers and executor contracts

Uploads call session.write_file(path, content). Its default implementation writes binary stdin with cat, creates parent directories, and requests no snapshot. Custom executors can override write_file without changing this adapter. Downloads call session.read_file(). Lazy sessions read the active VM through live inspect, so downloads preserve running processes and do not force a snapshot.

Sync sandbox	Async sandbox
`execute(command)`	`await aexecute(command)`
`upload_files([(path, bytes)])`	`await aupload_files([(path, bytes)])`
`download_files([path])`	`await adownload_files([path])`

File paths must be absolute. Transfers return per-file responses; inspect error before reading their content. The async adapter’s synchronous methods raise NotImplementedError. Each adapter serializes its own calls, but callers sharing an executor must coordinate other mutations themselves. To use another backend implementation, follow Customize the SDK. See deepagents sandboxes for API signatures.

Get Started

Inference

Observability

Post-training

Data Lab

Sandboxes

Teams & Access Management

Other Capabilities

Integrations

Platform Notices

Install and test the backend

Connect the backend to an agent

Advanced: file transfers and executor contracts

​Install and test the backend

​Connect the backend to an agent

​Advanced: file transfers and executor contracts

Install and test the backend

Connect the backend to an agent

Advanced: file transfers and executor contracts