Codex-maxxing for long-running work

Source: OpenAI Date: 2026-06-22 URL: https://openai.com/index/codex-maxxing-long-running-work

Summary

OpenAI whitepaper (by Jason Liu) describing “codex-maxxing” — a workflow pattern for using Codex as a persistent, steerable workspace across long-running projects rather than for single-shot completions. Core concepts: durable threads that accumulate context across sessions, explicit steering (correcting or queuing Codex mid-task), and decomposing large goals into discrete verifiable steps. Positioned as broadly applicable — software, research, operations — not just coding.

Implications

• Long-running / autonomous agents: This is a direct operational document for multi-session autonomous work. Durable threads + mid-task steering is the same design space as persistent agent loops; it signals that OpenAI is actively codifying how to run agents across time, not just single requests.
• Floor fencing / reversibility: Steering-as-concept is a legibility and reversibility mechanism — the human can redirect or halt between steps. Worth watching whether the actual Codex product enforces this or just describes it.
• Two clocks: Closed-frontier side of the clock. Codex-maxxing is a frontier-model workflow pattern; no open-weight equivalent has comparable task-persistence infrastructure yet. Widens the usability gap between closed and open.