LangChain's Managed Deep Agents v0.9 Adds Schedules and Per-Run Setup

Agents that schedule their own work

The new Schedules SDK allows agents to create reminders, follow-ups and recurring tasks in the middle of a conversation. LangChain gives "Remind me about this tomorrow" as an example of a request the agent can now handle. The agent creates the schedule itself, and the schedule runs as the person who asked, using that person's permissions and connections. Results return to the same channel.

The company recommends putting the feature behind a tool, so that users can create, list, update and delete schedules just by chatting. In its example, the agent passes a prompt and a cron expression, and each time the cron fires, LangSmith starts a new run with that prompt.

Schedules inherit the channel where they were created. A weekday reminder requested in Slack posts a new message to that conversation on every run. One-time schedules, which use at instead of cron, reply in the original thread. LangChain says this suits follow-ups such as "check on this deploy in an hour."

One deployment, configured per run

Agents can now be configured dynamically for each run. Developers set up the agent as a callable function that receives the runtime at the start of a run and returns a define_deep_agent definition. That definition chooses the model, instructions, skills, MCP servers and sandbox for that run.

According to LangChain, this removes the need to maintain a near-copy of the same agent for every team or repo. A single deployment can be customized based on context, such as the channel or user that originates the request. The company describes an internal Slack agent serving several teams: the function checks which channel a run came from and loads billing skills for finance, or an incident-response MCP server for the platform team.

LangChain argues this differs from telling the model in its instructions when to use a skill. The configuration is set before the model runs, so the agent never sees skills or MCP tools outside its configuration. The company says this keeps context small, works as access control, and lets developers pick the model itself, which the model cannot do. It also prevents non-deterministic outputs in cases where the exact set of tools is already known.

A second example covers a coding agent that loads different skills and a cheaper or more capable model for each repository. In the sample code, one repository maps to openai:gpt-5 and another to openai:gpt-5-mini. Callers choose the configuration by passing a context with the run, for example a repository name. The function falls back to a default when no context exists, such as on channel runs or when Agent Server loads the agent to read its schema.

Slack reactions and availability

• Acknowledgement: reactions show the sender right away that the agent picked up the message, even if it spends time reasoning and calling tools before replying.
• Defaults: reactions are on by default with the 👀 emoji. The new reactions option on a Slack channel can turn them off, pick another emoji, or choose one per message.
• Custom logic: in LangChain's example, a function returns 🐛 when a message mentions something broken and 👀 otherwise. A function can also call a model to pick a more precise emoji, and the company points to a decision model called Jev as a fast and cheap option.

Managed Deep Agents v0.9 is available in Public Beta today. Developers can start with the commands mda init and mda deploy and find more detail in the Managed Deep Agents docs. The previous release, v0.8, added per-user memory, custom HTTP channels for triggering runs from any service, and built-in web search powered by Parallel.

What it means for teams building internal agents

LangChain says the updates make it easier to build internal agents that live in Slack or in a team's own channels and work more like teammates. Self-created schedules let such agents follow up without a person re-prompting them. Per-run configuration lets one deployment serve different teams or repositories without duplicated agents. Reactions give users quick feedback that a request has been received.