Skip to main content
Managed Deep Agents stores deploy-owned instructions and skills, and optional durable memory, in LangSmith Context Hub. That split lets you change agent behavior without rebuilding application code, while the project remains the source of truth for lasting instruction and skill updates.
Managed Deep Agents is in public beta and available on LangSmith Cloud in the US region only.

​
What lives in Context Hub

Tools, middleware, MCP connectors, channels, schedules, sandboxes, and the agent definition ship with the compiled deployment. They are not synced to Context Hub. For ownership of managed runtime fields, see Agent definition.

​
Source of truth

The project and Context Hub both hold copies of instructions.md and skills/. Which copy wins depends on where you make lasting changes and how you resolve conflicts on deploy.
  • Project files: Treat instructions.md and skills/ in the project as the source of truth for lasting changes. Each mda deploy syncs those files into the agent’s Context Hub repo.
  • Context Hub edits: Edit files in Context Hub when you need a quick change that should apply without a code redeploy. Those Hub edits stay live until a later deploy overwrites them, or until you choose to keep the Hub copy when a conflict prompt appears.
  • Durable memory: Memory under /memories/agent/ is agent-owned. Deploys enable the tree but do not overwrite content already stored there.

​
How deploy syncs context

mda deploy syncs deploy-owned context as part of the deploy pipeline:
  • Instructions and skills: Each deploy updates the agent’s Context Hub repo from the project files. Later deploys sync the project copies again. For skills, deploys also remove deployed skill files that no longer exist locally.
  • Durable memory: When memory is enabled, the deployment mounts one Context Hub tree at /memories/agent/. Deploys do not overwrite content already stored there.
  • Conflicts: If instructions.md or skills/ changed in Context Hub since the last mda deploy, the CLI stops before syncing and asks whether to overwrite the Context Hub copy with the project copy. The default answer keeps the Context Hub version and skips syncing that section. In a non-interactive shell there is no prompt. The deploy exits with an error, and you must re-run with --context-strategy overwrite or --context-strategy keep-hub. Re-running mda deploy on its own reports the same conflict again.
  • Concurrent edits: If the repo changes during the sync itself, the deploy fails with a commit conflict. Re-run mda deploy.
Interactive conflict prompt:
Non-interactive conflict (CI or redirected output):
For the full deploy step list and flags, see the CLI reference. For secrets routing and deploy options, see Deploy an agent.

​
Edit context in LangSmith

After a successful deploy:
  1. Open the deployment in LangSmith.
  2. Use the Details sidebar link to the Context Hub repo that holds the deployment’s instructions, skills, and memories. Archived repos are labeled when the context is no longer live.
  3. Edit files in Context Hub when you need a quick change that should apply without a code redeploy.
For lasting changes, edit the project copies and redeploy. For how deploy resolves Hub edits, see Source of truth. For browsing commits, promoting environments, and Context Hub concepts outside Managed Deep Agents, see Use the Context Hub and Context engineering concepts.

​
Local development

mda dev creates a local Context Hub mock for instructions, skills, and memory. Local Studio sessions do not create or update the hosted Context Hub repo for a deployment. For more information, see Develop locally with LangSmith Studio.

​
Choose where context belongs

​
See also

Connect these docs to Claude, VSCode, and more via MCP for real-time answers.
Edit this page on GitHub or file an issue.