How to tell whether a create_deep_agent call is a working Deep Agent or one in name only
Reviewing a codebase to answer whether an agent was really a LangChain Deep Agent, the constructor call alone was misleading. One agent was built with deepagents.create_deep_agent but passed a model-call middleware that returns request.override(system_message=<own prompt>, messages=<own messages>, tools=[]). That override replaces the library’s base prompt and strips every tool on each call, so the agent has no file system, no task/subagent tool and no planning instructions: it is a single plain model reply that only happens to run inside the deepagents graph (the author’s comment said this was deliberate, to keep replies to a fixed short format). Checklist that settled it quickly: (1) grep for create_deep_agent( versus create_agent( per module; (2) read every middleware passed to it and look for request.override(tools=...) or a tool allowlist, since that decides what the model can actually call; (3) check whether subagents= is passed or the default general-purpose subagent is reachable; (4) check whether anything reads or writes the agent’s files state; (5) find who starts the other agents: if a graph node or server function launches them, the agent is not orchestrating regardless of its name. Describe the result by capability (delegates, uses files, gated by code) rather than by which constructor built it.