Codex as a persistent workspace for long-running AI workflows

The article tackles the tension between AI’s single-turn interaction model and the reality that real work—engineering, research, creative projects—unfolds across hours, days, or weeks. Current LLM interfaces reset context with every new conversation, making sustained progress fragile. The problem isn’t prompt engineering; it’s that the tooling treats complex projects as isolated queries rather than continuous workstreams.

The concrete path is treating Codex as a persistent workspace rather than a chatbot. Jason Liu advocates breaking ambitious goals into verifiable steps, each producing artifacts that feed into the next. Codex preserves context across these steps, letting you checkpoint progress, revisit decisions, and maintain thread coherence. Practical strategies include structuring prompts around output files rather than raw text, naming threads by project stage, and explicitly marking when human judgment overtakes automated execution. The operational insight is about delegation triage: let Codex handle deterministic, well-scoped tasks, but maintain oversight for ambiguous or high-stakes decisions.

For the serious builder, the takeaway is that long-running AI workflows demand infrastructure, not better prompts. Abandon the fantasy of a single transcendent prompt solving everything. Instead, invest in context continuity—a workspace model where the AI holds state across sessions, and the human monitors trajectories, not individual tokens. The article implicitly argues that the bottleneck in AI-augmented work isn’t model capability but tooling that respects the temporal structure of real projects.