How an OpenAI Engineer Turns Codex Into a 24/7 Autonomous Coding Agent

Let It Run Continuously

Jason Liu, author of the 13k-star Instructor library and recent OpenAI Codex hire, publishes a detailed guide ( Codex-maxxing ) on converting Codex from a single-turn assistant into a long-running work system. His core insight: most users close sessions after one Q&A, but Jason keeps cross-month giant threads alive — one for calendar, one for open-source projects, one for social monitoring — switched via Command-1 through Command-9. These threads accumulate months of context, preferences, and decisions, so the agent resumes work without re-briefing.

He dictates tasks instead of typing, arguing that speech preserves raw, branching, traceable intent without prompt engineering overhead. Combined with Codex's Steering feature, he can inject mid-execution corrections and walk away.

Heartbeats + @computer: The Automation Backbone

The Heartbeats + @computer combo transforms Codex into a scheduled worker. Heartbeats acts as a cron layer; Jason's "Chief of Staff" thread runs every 30 minutes: scans Slack and Gmail, prioritizes messages, drafts replies (human approves before send).

A more complex example: animation project reviews. He posts a video to a Slack review thread; Codex polls every 15 minutes. When teammates comment, Codex re-renders and posts the new version. Because the Slack MCP server lacks file upload, the agent uses @computer to click "Add file" and upload the artifact.

Another case: before a shower, he told Codex to watch an Amazon customer-service queue; when he returned, the refund had landed. The same pattern now extends to Google Docs comments, GitHub PR reviews — any feedback loop with a verifiable next step.

Verification Gates Replace Wishful Thinking

Jason stresses verification mechanisms as the only way to know when a task ends. He tasked Codex with porting Python's Rich library to Rust with a hard rule: all original Python unit tests must pass . Test pass/fail becomes the termination signal; failures trigger automatic repair cycles. As he puts it: "Ambition without verification is just a wish."

OpenAI's latest update graduates Goal mode from experiment to GA: give a final objective and acceptance criteria, and Codex drives autonomously for hours or days, with progress checks, course corrections, and pause/resume — provided the task has a clear, verifiable feedback loop.

Memory Lives in Your Files, Not the Platform

Jason rejects platform-locked memory. All persistent threads originate from a local Obsidian vault structured as TODO, people, projects, agent, notes. The root AGENTS.md encodes sync rules: personnel changes, project advances, completed todos must update the vault. He effectively bypasses Codex's built-in memory, storing canonical knowledge in version-controlled markdown — readable, diffable, rollbackable, portable across tools.

He acknowledges Codex's Chronicle (screen-capture context builder) but notes it remains an opt-in preview with permission, rate, and privacy limits; the filesystem remains the most reliable memory substrate.

Workspace Evolution & Reusable Workflows

Codex's sidebar now renders Markdown, tables, PDFs, PPTs. The embedded browser lets agents control pages via JavaScript; Jason simultaneously runs Storybook for UI review, Remotion Studio for animation, Slidev for decks. His favorite deliverable: a single index.html with embedded JS/CSS — zero deploy, instant run.

He packages successful flows as Connectors and Skills — reusable workflow templates so Codex need not relearn.

Recent remote capability: Codex continues after screen lock; mobile app allows live monitoring, approval, and takeover. "You clock out, it clocks in; you lock, it works — overtime KPI achieved."

Reference: https://x.com/jxnlco/status/2057153744630890620