The post says Codex now owns the OpenAI turn, including native thread state, tool continuation, compaction, code mode, and dynamic tool search. OpenClaw keeps ownership of channels, memory, sessions, cron, browser, media, and its own integration tools.
This should reduce prompt bloat and make OpenAI-backed agent turns feel more native. It also gives OpenClaw a cleaner model for separating product logic from the model loop.
If you're running OpenAI models in OpenClaw, authenticate through the guided path and keep the canonical model reference in config. The article suggests the best experience comes from letting Codex handle the low-level loop while OpenClaw handles everything around it.
Read Original Post →