Codex Maxxing: How OpenAI's Jason Liu Builds Persistent AI Workers

Let It Run Continuously

Jason Liu, author of the 13k-star Instructor library and new Codex team member, publishes "Codex-maxxing" — a playbook for converting Codex from a chat tool into a persistent work system. His core insight: most users treat Codex as single-turn Q&A, but he keeps cross-month "giant threads" alive for each workflow (calendar, open-source, social monitoring), accessible via Command-1 to Command-9. These threads accumulate months of history, preferences, and decisions, giving the agent continuity without re-contextualizing each session.

Jason prefers voice input over typing, arguing speech preserves raw, unstructured thought — including vague, jumping, or traceability-requiring ideas — without prompt engineering overhead. Combined with Codex's Steering feature, he can interject new instructions mid-execution and walk away.

Heartbeats + @computer: The Automation Combo

The key to making Codex act like an employee is Heartbeats + @computer . Heartbeats acts as a cron-like scheduler for the agent. Jason runs a "Chief of Staff" thread every 30 minutes: it scans Slack and Gmail, prioritizes messages, drafts replies (but does not send), and leaves final approval to him.

A more complex example: for an animation project, he posts a video to a Slack review thread and sets Codex to check every 15 minutes. When colleagues leave feedback, Codex re-renders a new version and replies in-thread. Because the Slack MCP server lacks file-upload support, the agent uses @computer to click "Add file" and upload the rendered output.

Another case: before showering, Jason told Codex to monitor an Amazon customer-service queue. When he returned, the refund had already been processed. Similar loops now extend to Google Docs comments, GitHub PR reviews — any feedback triggers the next step automatically.

Verification Mechanisms: The Stop Condition

Jason emphasizes that ambition without verification is just a wish. He tasked Codex with porting Python's Rich library to Rust with a hard requirement: all original Python unit tests must pass . Test pass/fail determines task completion; on failure, the agent keeps iterating. This closed-loop feedback is a prerequisite for the new Goal mode (now graduated from experimental), where users define a final goal and acceptance criteria, and Codex drives autonomously for hours or days, with progress checks, direction changes, and pause/resume.

Memory in Your Hands, Not the Platform's

Jason's second pillar: personal work memory must not live inside the platform . All long-term threads originate from an Obsidian vault structured as TODO, people, projects, agent, notes. A top-level AGENTS.md encodes rules: personnel updates, project progress, task closure must sync to the knowledge base. He effectively abandons Codex's built-in memory, storing core data in local, version-controlled files — readable, editable, diffable, and rollback-able. Rationale: as AI memory grows, locking it into one platform creates vendor risk; files are portable across tools.

He notes Codex's Chronicle (screen-capture context builder) remains an experimental preview with permission, rate, and privacy limitations — promising but not yet production-ready. The filesystem remains the most reliable memory infrastructure.

Upgraded Workbench & Reusable Workflows

Codex's sidebar now renders Markdown, tables, PDFs, PPTs directly. The built-in browser lets agents control pages via JavaScript; users can annotate side-by-side without window switching. Jason often runs Storybook (UI review), Remotion Studio (animation), and Slidev (slides) simultaneously in the panel. His favorite deliverable: a single index.html with embedded JS/CSS — zero deploy, zero server, runs on open.

He packages successful flows as Connectors and Skills — reusable workflow templates. Once a useful task is solved, the process is encapsulated; next time Codex invokes it directly without relearning.

Recent remote capability: Codex continues working after screen lock; mobile app allows real-time monitoring, approval, and takeover. "You clock out, it works overtime; you lock screen, it keeps going."