How a Top OpenAI Engineer Extracts Maximum Value from Codex

Let It Run By Itself

Jason Liu, author of the 13k‑star open‑source library Instructor and recent OpenAI Codex hire, published a detailed guide called “Codex‑maxxing” that exposes his entire Codex workflow.

The core idea is to convert Codex into a long‑running system that can survive across months. Instead of ending a session after a single Q&A, Jason keeps a set of “cross‑month persistent giant threads” alive, each dedicated to a specific workflow such as scheduling, open‑source project monitoring, or Slack scanning, and switches between them with Command‑1 to Command‑9.

These threads accumulate months of conversation history, preferences, and decisions, so when a task is resumed the agent already knows the context without needing a fresh prompt.

Jason issues tasks primarily by voice, arguing that speaking preserves the original thought process and avoids the need for painstaking prompt engineering.

When combined with Codex’s Steering feature, he can insert additional commands while the agent is executing, allowing him to “cut in line” and continue without waiting.

The decisive combination is Heartbeats + @computer . Heartbeats act as a periodic scheduler for the agent. For example, a “Chief of Staff” thread runs every 30 minutes, scans Slack and Gmail, drafts replies for messages that need attention, and waits for a human to approve sending.

In a more complex animation workflow, Jason posts a video to a Slack review thread, and Codex checks the thread every 15 minutes. If a colleague provides feedback, Codex re‑renders a new version and replies. Because Slack’s MCP server does not support file uploads, the agent uses @computer to click the “Add file” button and upload the rendered file.

Similar pipelines have been extended to Google Docs comments, GitHub PR reviews, and any scenario where feedback triggers the next step automatically.

A key safety net is a verification mechanism. Jason once asked Codex to migrate Python’s Rich library to Rust, imposing the hard requirement that all original unit tests must pass. If the tests fail, the agent continues fixing the code until the verification succeeds.

In the latest update, OpenAI promoted Goal mode from experimental to official. By specifying a final goal and acceptance criteria, Codex can autonomously pursue the task for hours or days, with the ability to check progress, adjust direction, or pause, provided the task has a clear, testable feedback loop.

Memory in Your Hands

Another central principle is that personal work memory should not be hosted inside the platform. Jason stores all long‑term thread data in an Obsidian vault organized into sections such as TODO, people, projects, agents, and notes. Changes are recorded in a top‑level AGENTS.md file, ensuring that knowledge is version‑controlled, manually editable, and can be rolled back if needed.

He therefore largely avoids Codex’s built‑in Chronicle memory, which captures screen content to build context but remains an experimental preview with limitations in permissions, rate limits, and privacy.

The Codex workspace UI has also been upgraded: the sidebar now renders Markdown, filters tables, and displays PDFs and PPTs. Agents can use the built‑in browser to control web pages with JavaScript, allowing users to annotate while viewing without switching windows.

Jason frequently opens Storybook for UI component review, Remotion Studio for animation, and Slidev for presentations side‑by‑side in the panel.

His preferred delivery format is a single index.html file containing JavaScript and CSS, requiring no deployment or server—just open the file and it runs.

Connectors and Skills are treated as reusable workflow templates; once a useful task is completed, it is packaged so the next time Codex can invoke it directly without relearning.

Recent remote capabilities let Codex continue working after the computer is locked, and a mobile client can monitor, approve, or take over tasks in real time, enabling “after‑hours” productivity.

Overall, the guide demonstrates how to transform Codex from a simple tool into an autonomous employee that can extend work hours, handle repetitive tasks, and free the human operator for higher‑level work.