LangChain mulls async-native deepagents: how should finished subagents notify the parent agent?

hwchase17 · x · 2026-10-13

LangChain founder Harrison Chase is exploring how to make deepagents more async-native, with async subagents as a first-class concept.

Two open design questions:

How does a finished subagent notify the parent? When the parent is still running, candidates are: (1) interrupt the parent run, apply the subagent update, then continue (like "steer"); or (2) wait for the parent run to finish and batch the update with other new messages (like "queue").
Human-in-the-loop handling: if the main agent is paused awaiting explicit human approval, should the subagent update wait and batch with the human response, or resume the thread (canceling the interrupt, possibly re-asking for the same approval)?

He's soliciting additional options from the community — a glimpse into the engineering challenges of concurrent multi-agent orchestration.

More from coding & agent