OpenAI Engineer's Codex Masterclass: Persistent Agents, Heartbeats & Local-First Memory

Let It Run Continuously

The core of Jason Liu's approach is reshaping Codex from a chat tool into a long-running work system. Most users close sessions after a single Q&A Jason keeps cross-month survival threads for each workflow — calendar, open-source projects, social monitoring — accessible via Command-1 through Command-9. These threads accumulate months of history, preferences, and decisions, so the agent resumes context automatically without re-briefing.

Jason prefers voice input over typing: speaking preserves raw, branching, traceable thought without prompt engineering. Combined with Codex's Steering feature, he can interject new instructions mid-execution and walk away.

Heartbeats + @computer: The Automation Combo

Heartbeats acts as a cron layer for the agent. Jason's "Chief of Staff" thread runs 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: during an animation project, he posts a video to a Slack review thread and sets Codex to poll every 15 minutes. When teammates comment, Codex re-renders a new version and replies in-thread. Because the Slack MCP server lacks file upload, the agent uses @computer to click "Add file" and upload the render itself.

Another case: before a shower, Jason told Codex to watch the Amazon support 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, Not Wishes

Jason insists on verification mechanisms to decide when a task ends. He once 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 sole completion signal; failures trigger automatic repair loops. As he puts it: "Ambition without verification is just a wish."

OpenAI has since graduated Goal mode from experiment: 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 Stays in Your Files

Jason rejects platform-locked memory. All long-lived threads originate from an Obsidian vault structured into TODO, people, projects, agent, notes. The top-level AGENTS.md encodes sync rules: personnel changes, project moves, task closures must update the vault. He effectively bypasses Codex's built-in memory, keeping core knowledge in local, version-controlled files that he can inspect, diff, roll back, and migrate freely.

Codex's own Chronicle (screen-capture context) remains an opt-in preview with permission, rate, and privacy limits — promising but not yet production-grade. For Jason, the filesystem remains the most reliable memory infrastructure.

Evolving Workbench & Reusable Workflows

The Codex sidebar now renders Markdown, tables, PDFs, PPTs, and embeds a browser controllable via JavaScript. Jason often runs Storybook, Remotion Studio, and Slidev side-by-side in the panel. His favorite deliverable: a single index.html with embedded JS/CSS — zero deploy, instant run.

He packages successful runs as Connectors and Skills — reusable workflow templates — so Codex can replay proven processes without relearning.

Recent updates add remote persistence : Codex continues working after screen lock, with mobile monitoring, approval, and takeover. The result: you clock out, it clocks in.

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