How an OpenAI Engineer Maximizes Codex: Persistent Threads, Voice Commands & Automated Workflows

Persistent Threads as Long-Running Work Systems

Jason Liu, author of the 13k-star Instructor library and new Codex team member, shares his Codex-maxxing playbook. The core insight: treat Codex not as a single-shot chat tool but as a continuous work system. He maintains cross-month giant threads for distinct workflows — calendar management, open-source projects, social-media monitoring — and switches between them with Command-1 through Command-9. Each thread accumulates months of conversation history, preferences, and decisions, so the agent resumes context automatically without re-briefing.

Voice Input and Mid-Task Steering

Jason dictates tasks rather than typing. He argues voice preserves raw, unstructured thought — including vague, jumping, or traceability-requiring ideas — without prompt engineering overhead. Combined with Codex's Steering feature, he can inject new instructions while the agent is working, then walk away without waiting.

Heartbeats + @computer: Scheduled Autonomous Execution

The Heartbeats mechanism 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. For an animation project, he set a 15-minute Heartbeat to watch a Slack review thread; when colleagues commented, Codex re-rendered the video and posted the new version. Because the Slack MCP server lacked file upload, the agent used @computer to click "Add file" and upload the render itself. Another example: before a shower, he tasked Codex to monitor an Amazon customer-service queue; the refund was processed by the time he returned.

Verification-Gated Task Completion

Jason insists on verification mechanisms to define when a task ends. He tasked Codex with porting Python's Rich library to Rust, with the hard requirement that all original Python unit tests must pass. Test success became the sole completion criterion; failures triggered automatic repair loops. As he puts it: "Ambition without verification is merely a wish."

Goal Mode: Native Autonomous Pursuit

OpenAI recently graduated Goal mode from experiment to stable. Users specify a final objective and acceptance criteria; Codex then drives the work autonomously for hours or days, allowing progress checks, course corrections, or pauses. The prerequisite is a clear, verifiable feedback loop.

Local-First Memory: Obsidian Vault over Platform Lock-In

Jason rejects storing long-term memory inside Codex. Instead, he roots every persistent thread in an Obsidian vault organized into TODO, people, projects, agent, notes. An AGENTS.md file at the root codifies rules: any change to personnel, project status, or completed todos must update the corresponding vault entry. This approach gives him full control — manual edits, version diffs, instant rollback — and makes the knowledge base portable across tools. He views Codex's built-in Chronicle (screen-capture context) as promising but still immature due to permission, rate-limit, and privacy constraints.

Upgraded Workbench and Reusable Workflow Templates

Codex's sidebar now renders Markdown, tables, PDFs, and slide decks directly. An embedded browser lets the agent control pages via JavaScript while the user annotates side-by-side. Jason frequently runs Storybook, Remotion Studio, and Slidev simultaneously in the panel. His preferred delivery artifact is a single index.html with embedded JS/CSS — zero deployment needed. He also packages successful workflows as Connectors and Skills for one-click reuse.

Remote Continuity After Lock Screen

Latest updates let Codex keep working after the host machine locks. Mobile apps provide real-time monitoring, approval, and even task takeover. The result: the agent works while the human rests.