Skip to main content
This tutorial builds a research assistant one capability at a time. Complete the quickstart first to scaffold a project, add API keys, and run mda dev locally. Then add a search tool, use durable memory, run the agent on a daily schedule, and deploy it to LangSmith. For a complete project that uses every primitive, see the example project.
Managed Deep Agents is in private beta, available on LangSmith Cloud in the US region only. Join the waitlist to request access.

​
Build the agent

1

Write the instructions

Replace instructions.md with the research assistant’s behavior. The instructions reference the tool you add next and the durable memory the runtime provides:
instructions.md
# Research assistant

You are a careful research assistant. Find sources, keep notes, and return
concise answers with citations.

## Behavior

- Use the `web_search` tool to find sources instead of guessing.
- Cite the sources you used.

## Memory

- Remember the user's topics of interest and preferred answer format.
- Never store secrets, API keys, or credentials in memory.
2

Add a search tool

Create a tools/ module with a search tool, then import it into the agent entry. This example returns a placeholder result, so it runs without an external API. Replace the body with a call to your search provider.
from langchain.tools import tool


@tool(parse_docstring=True)
def web_search(query: str) -> str:
    """Search the web for a query and return result snippets.

    Args:
        query: The search query.
    """
    # Replace this stub with a call to your search provider.
    return f"Top results for '{query}': ..."
Import the tool into the agent entry and pass it to the definition:
from managed_deepagents import define_deep_agent

from tools.search import web_search

agent = define_deep_agent(
    model="openai:gpt-5.5",
    tools=[web_search],
)
For more on authored tools, see Custom tools.
3

Run the agent locally

Install dependencies and start the local dev server:
uv sync
mda dev .
mda dev opens the agent in LangSmith Studio. Send a question and confirm the agent calls web_search and answers with the returned snippets.
4

Use durable memory

Managed memory is on by default, so you do not configure a backend. The runtime stores durable memory in Context Hub, and the agent reads and writes it during runs. Because the instructions tell the agent to remember topics of interest, tell it a preference in one turn (“I only care about open-source releases”), then start a new conversation and confirm it recalls the preference.To turn off managed memory, set disable_memory=True or disableMemory: true in the agent definition. For how memory persists, see How Managed Deep Agents work.
5

Schedule a daily digest

Add a schedules/ module so the agent runs on a cron cadence without a user message. This schedule runs every weekday at 8am Pacific:
from managed_deepagents import define_schedule

schedule = define_schedule(
    cron="0 8 * * 1-5",
    timezone="America/Los_Angeles",
    prompt="Summarize what you learned yesterday and list open questions.",
)
mda deploy reconciles this schedule into a LangSmith cron job after the deployment is live. For thread behavior and constraints, see Schedules.
6

Deploy the agent

Deploy the project to LangSmith:
mda deploy .
On success, the CLI prints the deployment dashboard URL. The deploy syncs the instructions to Context Hub, uploads the compiled project, and reconciles the daily schedule. For deploy flags and troubleshooting, see Deploy an agent and the CLI reference.
7

Inspect the run

Open the printed URL in LangSmith to inspect build status and revisions. Open traces to inspect the agent’s inputs, model calls, web_search calls, memory reads and writes, and final responses.

​
Next steps

Custom middleware

Add logging, retries, limits, and guardrails around model and tool calls.

Connect MCP tools

Load tools from remote MCP servers.

Example project

See a complete project that uses every primitive.

How it works

Understand compilation, the deploy lifecycle, and Context Hub.
Connect these docs to Claude, VSCode, and more via MCP for real-time answers.
Edit this page on GitHub or file an issue.