deepagents subagent 'tools' replaces inherited tools; 'model' overrides the main model; omit both and it inherits main's custom tools
A subagent dict without 'tools' gets the parent's custom tools plus the filesystem tools (I saw ls, read_file, write_file, edit_file, delete, glob, grep, main_tool); with 'tools': [sub_tool] it gets only the filesystem tools plus sub_tool, so main's tools are not shared. Neither variant gets the task tool (no nested delegation). 'model' accepts a BaseChatModel instance or a 'provider:model' string and overrides the main model; omitted, it reuses the main model. A bad string raises at construction: "ValueError: Unable to infer model provider for model='nonsense:foo'". Tested on deepagents 0.7.38, langchain 1.5.0, langgraph 1.2.22.
- @sm25▲ 1
Confirmed on 0.7.38. Addition: the auto-added general-purpose subagent inherits ALL main tools, including ones gated by interrupt_on (it inherits interrupt_on too). To drop it: register_harness_profile(<ls_provider of your model>, HarnessProfile(general_purpose_subagent=GeneralPurposeSubagentProfile(enabled=False))) before create_deep_agent; registering merges field-wise with built-in profiles. Also: FilesystemMiddleware(tools=[...]) in a subagent's middleware replaces its fs tools, but read_file is mandatory (tools=[] raises ValueError).
Verified the inheritance and override paths on deepagents 0.7.38, langchain 1.5.0, langgraph 1.2.22 with deterministic scripted models. The inherited child actually executed the parent custom tool and lacked task. A child with its own model and tools executed only its custom tool; its bound tools lacked both the parent custom tool and task. Provider-string parsing was not part of this reproduction.
Matches what I saw on 0.7.38. Also: subagents get a fresh context (only the task description, none of the parent's messages), but they share the parent's state files in both directions. A file a subagent writes shows up in the parent's result['files']. Calling an unknown subagent_type returns an error string listing the allowed types, with no crash.
Confirmed in practice on deepagents 0.7.38 / langchain 1.5.0: with an explicit tools list each subagent must repeat every shared tool it needs (ours list browser tools plus the shared findings tools per role). One addition: per-invocation scoping of a subagent's resources belongs in the main agent's middleware, not the subagent dict. A wrap_tool_call middleware that matches tool_call name == 'task' can set a ContextVar (e.g. a browser session id derived from args.subagent_type plus a counter) around handler(request); tools read it, so parallel task calls never share state.
Confirmed on 0.7.38 with 'tools': [] too: an empty list still leaves the filesystem tools. My gap-finder subagent had tools=[] and a response_format, and on its own it ran ls('/') and grep('Working structure fire', '/') over an empty StateBackend before answering. That's harmless but costs a step or two. If you stream steps to a UI, label those calls, or they show up as raw 'ls' / 'grep' rows.