Latest in AI
LangChain Brings Mid-Run Slack Questions to Managed Deep Agents
A new beta announcement highlights Slack buttons and forms that pause Managed Deep Agents for human input, with practical limits developers need to handle.
On October 20, 2026, LangChain product manager Nathan Drezner announced a beta workflow that lets Managed Deep Agents ask questions through buttons in a Slack thread, pause, and continue after the user responds. The focus is richer human input during a task, beyond simply approving or rejecting the agent's next action.
That distinction matters when an agent is missing a decision rather than permission. A support agent might need someone to choose an escalation path. A research agent might need a preferred scope before continuing. Making a user answer a structured question can be clearer than asking the model to interpret another free-form message in a busy conversation.
This is not the launch of Managed Deep Agents itself. LangChain opened the platform's public beta on August 7. It already provided a managed runtime around the Deep Agents harness, with developers controlling instructions, models and tools while LangSmith handled the surrounding infrastructure. The October announcement draws attention to how an agent can ask for human input inside that runtime.
A form tied to a paused task
The agent-owned interrupt documentation explains the mechanism. A developer's tool posts a Slack Block Kit message or modal and calls an interrupt. The thread pauses at a checkpoint. A button press or form submission returns as a channel event, and the SDK matches that answer to the interrupted run so the tool can continue.
Developers own the form and the logic that uses the answer. Managed Deep Agents carries the response back, avoiding a separate webhook service for that interaction. The feature requires a deployed Slack channel and managed-deepagents version 0.8.0 or later. The forms render in Slack; Studio shows the underlying interrupt value.
The same documentation identifies limits worth testing. Interrupted tools restart from the top when resumed, so posting the form must be guarded against repetition. Multiple pending interrupts need unambiguous correlation. Modal submissions can answer only one parked interrupt, dismissing a modal leaves the task paused, and repeated button presses are not deduplicated automatically.
Those details make the feature more useful to evaluate than a short demo alone. The happy path is a question, a click and a resumed task. A production workflow also needs an answer for an abandoned question, an outdated response or two people trying to answer at once. Teams should test those cases before placing consequential actions after a form submission.
Choosing the right kind of human input
LangChain's custom-tool documentation describes the separate option of pausing before selected tool calls. Developers configure those interruptions on the agent definition, and the managed runtime provides the durable state needed to resume. That remains useful when the actual question is whether a proposed tool call should proceed.
The Slack integration guide says platform-rendered human-in-the-loop requests in Slack support approval and rejection. Agent-owned forms give developers room for their own choices and fields instead. Slack access also still needs authorization, and a workspace may require administrator approval during deployment.
The practical gain is a more deliberate conversation between an agent and the person directing it. A well-designed question can expose the missing decision at the moment it matters. The remaining product work is deciding when to ask, showing enough context to answer, and handling the answer safely when the task resumes.
I'm Paras, and I help technical founders explain developer products through the workflows they make possible.
If you're launching an agent feature, I can help turn the implementation details and real limitations into a clear story for the people building with it.