deepagents — tool call path: reasoning vs. execution runtime

deepagents — tool call path: reasoning vs. execution runtime A workflow diagram generated by Archify. 01 / Agent reasoning EX / Approval (HITL) 03 / wrap_tool_call + tool 04 / Execution runtime Decide Execute Observe Model · tool_calls · Agent reasoning › Decide Model tool_calls Observation · ToolMessage · Agent reasoning › Observe Observation ToolMessage interrupt_on? · after_model · Approval (HITL) › Decide interrupt_on? after_model Rejected · no execution · Approval (HITL) › Execute Rejected no execution Mutation guard · same-path edits · wrap_tool_call + tool › Execute Mutation guard same-path edits Tool function · permission check · wrap_tool_call + tool › Execute Tool function permission check Offload check · > 20k tokens · wrap_tool_call + tool › Execute Offload check > 20k tokens Backend · files + exec · Execution runtime › Execute Backend files + exec Sandbox shell · BaseSandbox · Execution runtime › Observe Sandbox shell BaseSandbox allowed approve / edit tool_calls file op / exec message or stub reject result execute() full text Legend Agent logic Policy Context / trace Cloud service External system