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 ship it with mda deploy. 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, 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 application code, custom HTTP routes, authentication beyond a LangSmith key or Supabase, stronger isolation, maximum scalability, or a region other than US.

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_API_KEY (to deploy) and the provider key its model requires (ANTHROPIC_API_KEY, OPENAI_API_KEY, …). Uncomment the right provider line and tell the user to paste both.

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.
• A model provider API key.

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
  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/tasks/<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. TypeScript declarations also accept .tsx, .mts, and .cts.

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.

To route through LangSmith Gateway (rate limits, fallbacks, workspace-billed credits), scaffold with mda init <name> --gateway. Gateway model slugs use provider/model-name, not provider:model-name.

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; cross-user thread access returns 403.

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.

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.

Sandboxes

A sandbox gives the agent an isolated filesystem and shell. mda init scaffolds one; delete the sandbox/ directory to opt out, which is right for an agent that only needs its prompt, tools, and memory.

# sandbox/__init__.py
from managed_deepagents import define_sandbox

sandbox = define_sandbox(
    scope="thread",
    idle_ttl_seconds=600,
    default_timeout=600,
)

// sandbox/index.ts
import { defineSandbox } from "managed-deepagents";

export const sandbox = defineSandbox({
  scope: "thread",
  idleTtlSeconds: 600,
  defaultTimeout: 600,
});

scope="thread" (the default) creates one sandbox per durable thread. scope="agent" shares a single filesystem across threads — only use it for intentionally shared state, since threads can then read and modify each other's files. Set the creation source with template_name or snapshot_id, never both.

The agent works through ls, read_file, write_file, edit_file, glob, grep, and execute. Use instructions.md to say where it should work and what it must not touch. mda delete also deletes the managed sandboxes.

During mda dev, if the provider is unavailable the runtime falls back to a local temp directory and prints the path. That fallback is for development only — verify sandbox behavior in a dev deployment.

Schedules

One schedule per file under schedules/, each exporting a named schedule. The file name becomes the schedule name.

# schedules/daily_digest.py
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.",
)

Define exactly one of prompt (turned into a user message) or input (a structured LangGraph input). cron must be a standard five-field expression; without timezone, crons run UTC.

Schedules use ephemeral threads by default — a fresh thread per run, deleted afterward. Pass thread={"mode": "persistent", "id": "..."} only when runs should accumulate durable thread state. Set deliver_to to post results through a configured Slack channel.

Declarations are extracted at compile time without running your code: use literals and top-level literal constants only. No env vars, function calls, or **kwargs.

mda deploy reconciles schedules after the deployment is live — it deletes MDA-owned crons and recreates them from the current files, so deleting a file and redeploying removes the cron. --no-wait skips reconciliation entirely, so never use it when adding, changing, or removing schedules.

Channels

A channel connects the agent to an external messaging service: inbound events start runs, and responses go back to the same conversation. Slack is the only supported provider. One channel per file under channels/, each exporting a named channel.

# channels/slack.py
from managed_deepagents import channels

channel = channels.slack()

The file name sets the channel name and its inbound route — channels/slack.py receives events at POST /channels/slack/events. Names must be unique; never name a file channels/channel.py.

Channel-originated runs expose runtime.channel to tools and middleware, carrying the normalized event and conversation address plus methods to post and update messages. Ordinary HTTP runs and scheduled runs have no originating channel, so runtime.channel is absent.

Slack setup needs a project-root slack-app-manifest.json and SLACK_SIGNING_SECRET + SLACK_BOT_TOKEN in .env. Treat the manifest as the source of truth; files generated under .mda/ are build artifacts and must not be committed. runtime.channel never exposes the bot token.

A channel receives messages that start runs. It is not the same as giving the agent Slack tools for initiating operations — a project may want either or both.

Evals

MDA evals are Harbor evals. evals/tasks/ is the canonical dataset; author complete Harbor tasks there. mda evals does not introduce a separate format and does not run trials — it packages the agent for Harbor and prints a harbor run command.

mda evals init smoke      # optional starter under evals/scaffold/
mda evals compile .       # copies scaffolds into evals/tasks/, writes the handoff

evals/ is not included in the deployed build. Harbor needs Docker for its default environment, and does not read .env — the generated job config writes ${VAR} placeholders, so export the variables in the shell that runs Harbor. Verifiers write a numeric reward to /logs/verifier/reward.txt or metrics to /logs/verifier/reward.json. For deeper eval design, see [[eval-engineering]].

CLI reference

Key flags:

• init: --model SPEC, --instructions TEXT, --instructions-file PATH, --memory agent|none, --gateway, --no-sandbox
• build: --out OUT (defaults to <path>/.mda/build, emptied before each build)
• dev: --port, --hostname, --no-browser, --no-reload
• deploy: --name, --deployment-type dev|prod, --workspace-id, --no-wait
• logs: --name, --lines, --level, --follow / --no-follow, --workspace-id
• delete: --name, --workspace-id, --yes

mda init detects the language from the current directory (pyproject.toml → Python, package.json → TypeScript, both or neither → interactive prompt). mda dev requires uv for Python and resolves the LangGraph dev server itself.

mda delete is destructive and removes the deployment plus its LangSmith resources. Confirm with the user before running it, and never pass --yes unprompted — that flag exists to skip the confirmation you should be getting.

Deploy and Context Hub

Authentication resolves in order: LANGGRAPH_HOST_API_KEY, LANGSMITH_API_KEY, LANGCHAIN_API_KEY — read from the project .env first, then the shell. In an interactive terminal with no key found, mda deploy prompts and saves it to .env. Use --workspace-id or LANGSMITH_WORKSPACE_ID with an organization-scoped key.

mda deploy routes local inputs to different managed surfaces:

instructions.md + skills/**   -> Context Hub deploy-owned context
.env                          -> deploy auth + non-reserved hosted secrets (never archived)
project source                -> .mda/build archive -> hosted deployment
schedules/**                  -> LangSmith cron jobs, after the deployment is live

Non-reserved .env entries — provider keys, tool credentials, database URLs — are forwarded as hosted deployment secrets. Reserved platform variables (LANGSMITH_API_KEY, LANGGRAPH_HOST_API_KEY, LANGCHAIN_API_KEY, LANGSMITH_WORKSPACE_ID) authenticate the deploy and route it, but are never uploaded as user-managed secrets. Deploy fails before upload if the model's provider key is not available from .env, the shell, or LangSmith workspace secrets.

Context Hub holds /instructions.md and /skills/** (deploy-owned, resynced each deploy) and /memories/agent/** (runtime-owned, preserved across deploys).

Troubleshooting: no agent entry file found → add agent.py at the root. 401/403 → the key's workspace lacks beta access. Context Hub conflict → re-run the deploy. Build over 200 MB → remove generated artifacts. BUILD_FAILED / DEPLOY_FAILED → open the printed URL and read the revision logs.

Human-in-the-loop

Pause before sensitive tool calls with interrupt_on, and gate filesystem paths with permissions:

agent = define_deep_agent(
    name="support-agent",
    model="anthropic:claude-sonnet-4-6",
    tools=[refund_customer],
    interrupt_on={"refund_customer": True},
)

interrupt_on applies the same behavior as LangChain's human-in-the-loop middleware; see [[langgraph-human-in-the-loop]] for approve/edit/reject semantics. Interrupts need durable thread state, and the managed runtime owns the checkpointer, so no extra setup is required.

Respond to interrupts in Studio during mda dev. On a deployed agent, resume through the LangGraph server API with a Command(resume=...) payload — but note that programmatic invocation from your own application is not documented during public beta.

Gotchas

• name= is required in define_deep_agent / defineDeepAgent. A definition without it fails.
• Model IDs need the provider prefix: anthropic:claude-sonnet-4-6, not a bare model name. Python uses google_genai:, TypeScript uses google-genai:, and Gateway uses provider/model.
• Do not set managed fields (backend, store, checkpointer, memory, skills, system prompt) in the agent definition.
• Memory is opt-in via memory.py, not a constructor argument. disable_memory is legacy — declare or delete memory.py instead.
• MCP connectors do not exist. connectors/mcp.* and define_mcp_servers were removed; writing them fails.
• Restart mda dev after adding a managed file. New memory.py, identity.py, schedules/, or channels/ declarations are discovered at compile time, not by hot reload.
• --no-wait skips schedule reconciliation and exits before DEPLOYED.
• Schedule declarations must be static literals — the compiler extracts them without running your code.
• .env is never archived, and .gitignore must keep it out of version control. Do not write live keys into it on a user's behalf.
• The docs run slightly ahead of the released CLI. Verify against mda --help and the installed package before trusting a flag or import. As of mda 0.5.0: the sandbox docs show sandboxes.langsmith(...), but that import raises ImportError — use define_sandbox(...) as shown above; and the documented mda init --identity and mda deploy --configure-slack flags are not present (identity.py is scaffolded by default).