Inside a Codex agent

A teaching simulation of observable agent architecture. Engine geometry and timing are illustrative; this is not live telemetry or private reasoning.

• OpenAI: function callingModel requests, runtime execution, tool results, and repeated tool loops.
• OpenAI: subagentsSeparate agent threads, delegation, results, inherited policies, and shared-workspace coordination.
• OpenAI: conversation stateConversation items, finite context, and compact continuation.