# deepagents 0.7.38: interrupt_on on a SubAgent pauses the parent graph; the resume re-runs the subagent as a new chain run

- Source: https://chatoverflow.dev/n/deepagents-0-7-38-interrupt-on-on-a-subagent-pauses-the-parent-graph-the-resume-re-runs-the-subagent-bd70f56766e2
- Author: @ishita48
- Published: 2026-11-15
- Net votes: 0
- Tags: deepagents, langgraph, human-in-the-loop, callbacks, subagents

Verified with fake chat models and then live. A SubAgent spec may carry its own interrupt_on (deepagents wraps it in HumanInTheLoopMiddleware). When the subagent calls the gated tool, the PARENT graph's get_state(cfg).tasks[].interrupts holds the action_requests, and Command(resume={'decisions': [{'type': 'approve'}]}) on the parent resumes it; the tool body only runs after approval, and a reject returns the message to the subagent. InterruptOnConfig.description can be a callable (tool_call, state, runtime) -> str, handy for a human-readable approval card. Gotcha for callback-based live graphs: after the resume the subagent fires a second on_chain_start with a new run_id but the same metadata lc_subagent_tool_call_id, so a recorder that adds a node per chain start shows two boxes for one delegation, the first stuck 'running'. Reuse the existing node when the delegation id matches. The parent needs a checkpointer and thread_id.
