Major update from @LangChain: Managed Deep Agents v0.9. It's their hosted runtime for long-running agents, and this release is aimed at internal agents that behave like teammates in Slack and other channels.

You now get:

Agents that schedule their own reminders, follow-ups and recurring tasks mid-conversation
One deployment that picks its model, skills, tools and sandbox fresh on every run
Slack reactions, so people know the agent saw their message before it replies

What's new?

Agent-created schedules: A new Schedules SDK lets agents handle requests like "remind me about this tomorrow" or "check on this deploy in an hour." The agent creates the schedule itself during the conversation, it runs as the person who asked using their permissions and connections, and results post back to the same channel. Recurring schedules use cron, and one-time ones reply in the original thread.
Per-run configuration: Define your agent as a function that receives the runtime at the start of each run and returns its setup: model, instructions, skills, MCP servers and sandbox. Instead of keeping near-copies of one agent for every team or repo, you run one deployment. The agent never sees skills or tools outside its configuration, which keeps its context small and doubles as access control. It also means the model isn't left to choose its own toolkit when you already know what it should have.
Slack reactions: The agent reacts to the message that starts a run while it's still reasoning and calling tools. You can turn it off, pick another emoji, or choose one per message with a function. You can even use a model to pick a more fitting emoji, like 🐛 for something broken.

You now get:

Agents that schedule their own reminders, follow-ups and recurring tasks mid-conversation
One deployment that picks its model, skills, tools and sandbox fresh on every run
Slack reactions, so people know the agent saw their message before it replies

What's new?

Agent-created schedules: A new Schedules SDK lets agents handle requests like "remind me about this tomorrow" or "check on this deploy in an hour." The agent creates the schedule itself during the conversation, it runs as the person who asked using their permissions and connections, and results post back to the same channel. Recurring schedules use cron, and one-time ones reply in the original thread.
Per-run configuration: Define your agent as a function that receives the runtime at the start of each run and returns its setup: model, instructions, skills, MCP servers and sandbox. Instead of keeping near-copies of one agent for every team or repo, you run one deployment. The agent never sees skills or tools outside its configuration, which keeps its context small and doubles as access control. It also means the model isn't left to choose its own toolkit when you already know what it should have.
Slack reactions: The agent reacts to the message that starts a run while it's still reasoning and calling tools. You can turn it off, pick another emoji, or choose one per message with a function. You can even use a model to pick a more fitting emoji, like 🐛 for something broken.

Managed Deep Agents v0.9

Agents that schedule their own follow-ups

Agents that schedule their own follow-ups