OpenAI's Codex team is focusing on improving how its models work inside OpenClaw, a popular open-source tool for running AI agents. Codex engineering lead Tibo Sottiaux says the team is working with OpenClaw creator Peter Steinberger and promises more updates next week.
https://twitter.com/thsottiaux/status/2047000994354241562
Codex product lead Nik Pash discovered the core issue while debugging. When OpenClaw was configured to use the Codex harness with OpenAI models, an authentication bug would cause the system to silently fall back to the Pi harness. Users thought the Codex harness was running, but it wasn't.
Pash submitted two pull requests to fix the authentication bridge and prevent the silent fallback.
The difference in agent behavior before and after the fix is stark. With the Pi harness, the agent would only do shallow polling on each heartbeat: read a heartbeat file, check Discord, return a HEARTBEAT_OK status, and ignore other instructions. Sometimes it would figure out what to do but never actually make the tool call.
Switch to the Codex harness, and the agent enters a full work cycle: reading workspace context, parsing task lists, checking repositories, executing edits, and attempting verification. Subsequent heartbeats can pick up where the last one left off instead of repeating work.
Pash says the agent's prompts and higher-level workflow logic weren't changed at all. The improvement came entirely from swapping the underlying runtime adapter—the harness—which changed how OpenClaw communicates with the model API. This fix arrives as OpenAI's developer tool continues to grow, recently hitting 4 million monthly active users in under two weeks.