LangChain's async subagents and Perplexity's Teammates both move agent work to the background. Design for status, not waiting

When an agent works in the background, the hard part stops being the model and becomes the job queue around it.

What happened

LangChain added async subagents to Deep Agents. An orchestrator stays available to the user while subagents run long tasks in the background and report back. The orchestrator gets five tools: start, check, update, cancel and list tasks. When it sends a running task new instructions, the subagent's update strategy decides what happens: "interrupt", the default, stops and restarts it with the new instruction, while "enqueue" holds the follow up until the current run finishes. Unfinished tasks are listed in the orchestrator's system prompt, and a subagent waiting for approval shows up as waiting on its own thread. LangChain notes task notifications are in beta and the API may still change. The post says Claude Code and OpenAI's dots already push work to background workers.

The same day, The Verge reported that Perplexity launched Teammates, always on work agents that can take different roles in a workspace and run tasks at the same time. Each Teammate gets its own thread in the Perplexity app, and you can interact with them in Slack. It requires the $200 per month Max plan.

My take

This is how good automation has always worked. In n8n or Make, you answer the webhook fast and push the slow work to a queue. Agents are learning the same lesson.

If you build long running agents for clients, give every background job the controls you would give any queue:

1. A visible status: queued, running, waiting for approval, done or failed.
2. A cancel button that actually stops the work and any spend.
3. A clear rule for new instructions: restart now, like interrupt, or after the current run, like enqueue.
4. Approvals that reach the right person in Slack or email without blocking everything else.
5. A timeout and a cost cap per task.

The interrupt versus enqueue choice is the one I see skipped. A client who says "use the new price list" wants an interrupt. One who says "also send the report" wants it queued. Decide that up front, or the agent will guess.