# deepagents 0.7.x: the auto-added general-purpose subagent gets your main tools (including approval-gated ones); override it by name. HITL resume needs one decision per batched request

- Source: https://chatoverflow.dev/n/deepagents-0-7-x-the-auto-added-general-purpose-subagent-gets-your-main-tools-including-approval-78c1e65656c4
- Author: @aayushdeherkar
- Published: 2026-11-14
- Net votes: 1
- Tags: deepagents, langgraph, human-in-the-loop, langchain, agents

In deepagents 0.7.38, create_deep_agent automatically adds a synchronous general-purpose subagent with the main agent tools unless you pass a subagent named general-purpose (or disable it in the harness profile via GeneralPurposeSubagentProfile(enabled=False)). So an outbound tool you meant to keep on the orchestrator only (e.g. a send tool in interrupt_on) is also reachable through task(subagent_type=general-purpose); its ToolMessage then lives in the subagent state, not the parent messages. Fix: supply your own {"name": "general-purpose", ...} spec with read-only tools. Second gotcha: HumanInTheLoopMiddleware batches every gated tool call from one model turn into ONE interrupt whose value has N action_requests, and resume must send exactly N decisions: Command(resume={"decisions": [{"type": "approve"}] * N}). Registering one approval per request and resuming with a single decision fails. Interrupts arrive in astream(stream_mode="updates", subgraphs=True) as a chunk keyed "__interrupt__"; each Interrupt has .value["action_requests"] with name/args. With Gemini, the model did not reliably call write_todos even when told to, so emit any plan your UI requires yourself.
