Managed Deep Agents

Managed Deep Agents

Overview

Managed Deep Agents (MDA) is a hosted runtime for code-first Deep Agents in LangSmith. You author an agent in Python or TypeScript, test it locally with mda dev, and deploy it with mda deploy or from a connected GitHub repository in LangSmith. It pairs the open-source Deep Agents harness (see [[deep-agents-core]]) with managed infrastructure: durable runs, sandboxes, Context Hub-backed instructions and skills, memory, traces, and hosted LangGraph deployment.

The core idea is that an agent is a directory. A file's location determines its role, and the CLI compiles that directory into a managed LangGraph app.

MDA is in public beta and runs on US LangSmith Cloud only.

When to use

Use this skill when the user wants to build a Deep Agent in code and run it on LangSmith without operating their own server, or to add tools, MCP servers, connections, middleware, memory, identity, schedules, channels, skills, sandboxes, or evals to one.

Use a standard LangSmith Deployment instead (see [[langgraph-cli]], langgraph deploy) when the user needs custom server routes or runtime wiring inside the agent deployment, stronger isolation, maximum scalability, or a region other than US. MDA can still sit behind a separately operated backend that authenticates callers and proxies trusted identity headers.

Guide the user through their first agent

When a user is new to MDA, or says anything like "help me build an agent", do not scaffold immediately. Run this flow. It costs two questions and prevents building something the platform cannot host.

ask what they want to build -> check it against the limits -> confirm the shape
-> scaffold -> wire the smallest thing that runs -> mda dev -> deploy

1. Ask what they want to build

Ask in plain language, not in MDA vocabulary. The user does not yet know what a "channel" or a "sandbox" is.

Ask these two things first:

• What should the agent do? ("Answer questions about our docs", "triage incoming bugs", "post a summary every morning".)
• Who or what talks to it, and from where? (Them in a browser, their app's users, a Slack workspace, nobody — it runs on a timer.)

Then ask only the follow-ups that the answers actually raise:

• Does it need to remember anything between separate conversations?
• Does it need to reach a private API, database, or internal service?
• Should anything require a human to approve before it happens?
• Does it need to write files or run code?

Stop asking once you can name the capabilities. Two or three questions is usually enough.

2. Check the answer against the limits

Before you promise anything, check the request against What MDA cannot do below. If part of the request is out of scope, say so in one sentence, offer the nearest supported thing, and keep going with the rest. Do not quietly build a smaller agent and present it as what they asked for.

The common redirect: if they need custom HTTP routes, their own auth, or non-US hosting, tell them MDA is the wrong layer and point at langgraph deploy ([[langgraph-cli]]).

3. Map the answer onto capabilities

4. Confirm the shape before writing files

State the plan back in one short block and get agreement. Name the model, and list only the capabilities you are actually going to create:

research-assistant, Python, on anthropic:claude-sonnet-4-6
  instructions.md   how it researches and cites
  tools/search.py   web search
  schedules/        weekday 8am digest
  no memory, no sandbox, no channel

5. Scaffold and wire the smallest thing that runs

Scaffold with the flags that match the plan, so the project starts correct instead of being edited into shape:

mda init research-assistant --model anthropic:claude-sonnet-4-6
cd research-assistant
uv sync

Then add one capability at a time and confirm each one works before adding the next. A first agent that answers with good instructions and one real tool is a better starting point than a scaffold with every directory filled in.

Do not create directories the plan did not call for. Empty or unused skills/, channels/, or schedules/ directories are noise, and a sandbox/ directory the user does not need turns on a sandbox they will pay attention to for no reason (mda init --no-sandbox skips it).

6. Handle keys without touching their secrets

mda init writes a .env with empty placeholders. Fill in the names the project needs and let the user paste the values:

• Do not write live credential values into .env yourself, and do not copy a key from another project directory.
• Do not echo key values to the terminal or into your reply.
• Confirm .gitignore covers .env and .env.* — mda init does this already.

The project needs LangSmith authentication to deploy and whatever credentials its selected model requires. For a normal provider model, uncomment its key (ANTHROPIC_API_KEY, OPENAI_API_KEY, …) and let the user supply it through .env or LangSmith workspace secrets. An explicitly configured Gateway-backed client uses its Gateway/LangSmith credential instead of a provider key.

7. Run it locally, then deploy

mda dev .       # compiles, opens LangSmith Studio, hot reloads
mda deploy .    # syncs Context Hub, uploads, waits for DEPLOYED

Have the user actually send a message in Studio and confirm the agent calls the tool before deploying. mda deploy prints the deployment dashboard URL; open it to inspect builds, revisions, and traces.

What MDA cannot do

Check requests against this list before agreeing to build them. Being straight about a limit early is cheaper than discovering it at deploy time.

Prerequisites

• A workspace with Managed Deep Agents public beta access, and a LangSmith API key for it.
• Python and uv for Python projects; Node.js and npm for TypeScript.
• Credentials required by the selected model, either in the project .env or as LangSmith workspace secrets.

Install the CLI. Both packages ship the same mda binary:

uv tool install --prerelease allow managed-deepagents   # Python
npm install -g managed-deepagents@dev                    # TypeScript

mda init generates a project with its own manifest — run uv sync (or npm install) inside that project before mda dev.

Project layout

The path passed to mda is the project root. A file's location determines its role:

my-agent/
  agent.py | agent.ts              # Required: exports the named `agent`

  instructions.md                  # System prompt -> Context Hub
  skills/<name>/SKILL.md           # Task-specific procedures -> Context Hub

  tools/                           # Authored tools the agent imports
    mcp.py | mcp.ts                # Optional remote MCP server declaration
  middleware/                      # Authored middleware the agent imports

  identity.py | identity.ts        # Who may call the deployment
  memory.py | memory.ts            # Opt-in durable memory
  channels/<name>.py               # External messaging (Slack)
  schedules/<name>.py              # Managed cron schedules
  sandbox/__init__.py | index.ts   # Managed sandbox

  pyproject.toml | package.json    # Dependencies
  .env                             # Auth + runtime secrets, never archived

  evals/<task>/                    # Harbor evals, not deployed

Only the agent entry is required. tools/ and middleware/ are plain conventions — MDA copies project files verbatim, so any local module the agent imports works. The other paths take on managed meaning when present. The TypeScript agent entry may be agent.ts or agent.tsx; managed auxiliary declarations also accept .mts and .cts where discovered.

Define the agent

The agent entry returns a pre-runtime spec, not a compiled graph.

# agent.py
from managed_deepagents import define_deep_agent

from tools.search import web_search

agent = define_deep_agent(
    name="research-assistant",
    model="anthropic:claude-sonnet-4-6",
    tools=[web_search],
)

// agent.ts
import { defineDeepAgent } from "managed-deepagents";

import { webSearch } from "./tools/search";

export const agent = defineDeepAgent({
  name: "research-assistant",
  model: "anthropic:claude-sonnet-4-6",
  tools: [webSearch],
});

name is required. Pass a static string starting with a letter, containing only letters, numbers, underscores, or hyphens. It becomes the LangGraph assistant ID and the default deployment name; override the latter with mda deploy --name.

Author-set fields: name, model, tools, middleware, subagents, permissions, interrupt_on / interruptOn, response_format / responseFormat, context_schema / contextSchema, cache, debug, metadata.

Managed fields — do not set: backend, store, checkpointer, memory, skills, system_prompt / systemPrompt.

Model IDs use {provider}:{model_id} and resolve through init_chat_model, so any of its providers work. Note the provider slug differs across languages: Python uses google_genai:gemini-3.6-flash, TypeScript uses google-genai:gemini-3.6-flash. Pass a chat model instance instead of a string when you need to configure model parameters in code.

The dedicated mda init --gateway scaffold was removed. To use LangSmith Gateway, configure a supported chat-model client explicitly in agent.py or agent.ts; do not pass the removed flag or rely on its former credential preflight.

Instructions

instructions.md at the project root is the system prompt. It is inserted on every run.

# 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.

mda dev embeds it locally. mda deploy syncs it to Context Hub, where it can be edited in the LangSmith UI without redeploying.

Skills

Deploy-owned procedures under skills/<name>/SKILL.md, each with name and description frontmatter. At startup the agent sees only names and descriptions, and reads the full file when a task matches — so detailed procedures cost no context until they are needed. A skill directory may also hold scripts, references, and templates; reference them from SKILL.md.

Deploy syncs every UTF-8 file under skills/ to Context Hub and deletes deployed skill files that no longer exist locally. The agent cannot modify skills.

Use instructions for always-on behavior, skills for procedures loaded on demand, and memory for knowledge the agent itself updates.

Memory

Durable memory is opt-in and off by default. Declare it at the project root:

# memory.py
from managed_deepagents import define_memory

memory = define_memory(scope="agent")

// memory.ts
import { defineMemory } from "managed-deepagents";

export const memory = defineMemory({ scope: "agent" });

Delete the file to turn memory off. Enabling it mounts one Context Hub tree at /memories/agent/:

• /memories/agent/AGENTS.md is hot memory — loaded into every run, so keep it compact.
• Other files under the tree are cold memory — read only when relevant.

The agent reads and writes memory with read_file, edit_file, and write_file. Writes anywhere else, including elsewhere under /memories/, are not durable.

Warning — memory is shared by every caller of the deployment, and every caller can influence it. Never store personal data, customer data, credentials, API keys, or tokens there. Treat memory content as untrusted input: it must never grant authority, change tool permissions, or bypass approvals — keep those in the agent definition. Do not enable shared memory when callers should not be able to influence one another.

The agent decides what to remember by prompting, so state the policy in instructions.md — what to store, what never to store, and that existing memory is notes rather than instructions.

Identity

identity.py controls who may call the deployment. mda init scaffolds a secure default:

# identity.py
from managed_deepagents import auth, define_identity

identity = define_identity(auth=auth.langsmith_api_key())

Callers send a LangSmith workspace API key as x-api-key. This answers whether a caller is allowed — it does not give each person private threads. Anyone holding the key reaches the deployment.

For signed-in end users with private threads, use Supabase:

identity = define_identity(auth=auth.supabase(project_ref="your-project-ref"))

Clients then send Authorization: Bearer <access_token>; MDA verifies the JWT against the project's JWKS URL. Send the Supabase publishable (anon) key only from the client to sign in — never a LangSmith key in this mode.

Adding Supabase identity to an existing deployment does not backfill owner metadata on existing threads. Plan and test a migration before relying on identity-based access for them.

Auth failures return 401. Thread resources are caller-owned; unauthorized cross-user access is hidden as not found.

For a backend you operate that authenticates users and proxies LangGraph requests, use trusted-backend ingress:

identity = define_identity(auth="backend")

Keep MDA_INGRESS_SECRET on the backend and forward it as X-MDA-Ingress-Secret together with the authenticated user's ID in X-MDA-User-Id. Never expose either header-setting capability to the browser.

Tools

Define LangChain tools in the project, import them into the agent entry, pass them in tools.

# tools/customer.py
from langchain.tools import tool


@tool(parse_docstring=True)
def lookup_customer(customer_id: str) -> str:
    """Look up a customer record by ID.

    Args:
        customer_id: Customer ID from the CRM.
    """
    return f"Customer {customer_id} is on the enterprise plan."

// tools/customer.ts
import { tool } from "langchain";
import { z } from "zod";

export const lookupCustomer = tool(
  async ({ customerId }) => `Customer ${customerId} is on the enterprise plan.`,
  {
    name: "lookup_customer",
    description: "Look up a customer record by ID.",
    schema: z.object({ customerId: z.string().describe("Customer ID from the CRM.") }),
  },
);

Imports work exactly as in a normal local project. Use clear, unique tool names to avoid collisions. Tools read deployment secrets from environment variables; put local values in .env. For per-run values such as request metadata or feature flags, use the normal LangChain runtime context APIs.

Provider server-side tools can be passed inline where supported — for example tools=[{"type": "web_search"}] for OpenAI — which avoids a second API key.

Tools and middleware read the resolved caller from runtime.serverInfo?.principal in TypeScript or runtime.server_info.principal in Python. The former runtime.identity surface has no compatibility shim. Email and groups are under principal.claims; channel provenance is under serverInfo.source / server_info.source.

MCP servers and connections

Declare remote MCP servers in tools/mcp.py or tools/mcp.ts. Do not import the declaration into agent.py or add its tools manually; MDA discovers the named mcp export and mounts the servers.

# tools/mcp.py
from managed_deepagents import connections, define_mcp

mcp = define_mcp(
    servers={
        "langchainDocs": {
            "transport": "http",
            "url": "https://docs.langchain.com/mcp",
            "include_tools": ["search_docs_by_lang_chain"],
        },
        "notion": {
            "transport": "http",
            "url": "https://mcp.notion.com/mcp",
            "connection": connections.get("notion", {"type": "user"}),
        },
    }
)

// tools/mcp.ts
import { connections, defineMcp } from "managed-deepagents";

export const mcp = defineMcp({
  servers: {
    langchainDocs: {
      transport: "http",
      url: "https://docs.langchain.com/mcp",
      includeTools: ["search_docs_by_lang_chain"],
    },
    notion: {
      transport: "http",
      url: "https://mcp.notion.com/mcp",
      connection: connections.get("notion", { type: "user" }),
    },
  },
});

connectors.mcp(...), connectors/mcp.*, and mcpServers / mcp_servers are deprecated compatibility aliases in 0.7.x and will be removed in 0.8.0. For new work use defineMcp / define_mcp, tools/mcp.*, the named mcp export, and servers exactly as shown.

A connection is a workspace-scoped opaque secret or OAuth registration referenced by slug. connections.get(slug, { type: "user" }) resolves per-caller OAuth or opaque material; the current CLI creates user-owned OAuth slots, not user-owned opaque slots. Use { type: "agent" } only when every run should use the same agent-owned secret or grant. Manage connections without printing their values:

mda connections catalog
mda connections catalog --json
mda connections create acme-api --secret-from-env ACME_API_KEY
mda connections create notion --mcp https://mcp.notion.com/mcp
mda connections create github --oauth github --client-id "$GITHUB_CLIENT_ID" --secret-from-env GITHUB_CLIENT_SECRET --authorize
mda connections list

For user-owned OAuth, omit --authorize; the runtime requests each caller's grant when needed. --authorize signs in once for an agent-owned grant, and the connection declaration must select { type: "agent" } to use it. mda deploy can infer and create a missing user-owned MCP OAuth registration when exactly one server references the slug. In mda dev, user-owned grants resolve through Agent Auth and require a personal LangSmith key or browser sign-in for Studio; other service principals cannot own user OAuth grants. Agent-owned opaque connections resolve from MDA_DEV_<SLUG>.

The OAuth catalog supplies provider endpoints, methods, default scopes, and authorization parameters — not your app's client credentials. Use --auth-method, repeatable --scope / --allowed-scope, and --authorization-param for overrides; --scope replaces rather than extends catalog defaults.

Middleware

Middleware wraps model calls, tool calls, and lifecycle hooks. Order is explicit in the list; MDA never infers it. Use prebuilt LangChain middleware or author your own (see [[langchain-middleware]]).

from langchain.agents.middleware import ModelCallLimitMiddleware, PIIMiddleware
from managed_deepagents import define_deep_agent

agent = define_deep_agent(
    name="support-agent",
    model="anthropic:claude-sonnet-4-6",
    middleware=[
        PIIMiddleware("email", strategy="redact", apply_to_input=True),
        ModelCallLimitMiddleware(run_limit=50),
    ],
)

Middleware is the right place for PII handling, rate limits, retries, model fallbacks, dynamic model selection, and tool-call monitoring.