How an OpenAI Engineer Turns Codex Into a Persistent Autonomous Agent

Persistent Threads as Continuous Work Context

Jason Liu keeps multiple cross-month Codex threads alive — one for calendar, one for open-source projects, one for social monitoring — switching via Command-1 through Command-9. Each thread accumulates months of conversation history, preferences, and decisions, so the agent resumes with full context without re-briefing.

Voice Input + Steering for Low-Friction Delegation

Liu dictates tasks instead of typing, preserving raw, unstructured thought including traceability requirements. Combined with Codex's Steering feature, he can inject mid-execution corrections without waiting for the current step to finish.

Heartbeats + @computer: Scheduled Autonomous Execution

Heartbeats acts as a cron layer. A "Chief of Staff" thread runs every 30 minutes: scans Slack and Gmail, prioritizes messages, drafts replies (human approval required before send). For an animation project, a thread polls a Slack review channel every 15 minutes; on feedback, 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. In another case, Liu tasked Codex to monitor an Amazon support queue before a shower; the refund was processed by the time he returned.

Verification-Gated Loops Replace Wishful Prompting

Liu stresses that ambition without verification is just a wish. He tasked Codex with porting Python's Rich library to Rust with a hard gate: all original Python unit tests must pass. Test results determine task completion; failures trigger automatic repair cycles.

Goal Mode Graduation

OpenAI has now promoted Goal mode from experimental to stable. Users define a final objective and acceptance criteria; Codex drives the work autonomously for hours or days, with progress inspection, course correction, and pause/resume. The prerequisite is a clear, verifiable feedback loop.

Local-First Memory via Obsidian Vault

Liu rejects platform-locked memory. All long-term context lives in an Obsidian vault (TODO, people, projects, agent, notes). An AGENTS.md at the root codifies sync rules: personnel changes, project updates, task closures must update the vault. This enables manual editing, git diff history, and instant rollback. He views Codex's built-in Chronicle (screen-capture context) as promising but immature on permissions, rate limits, and privacy.

Upgraded Workspace: Sidebar Rendering, Browser Control, Reusable Skills

The Codex sidebar now renders Markdown, filterable tables, PDFs, and PPTs. An embedded browser lets agents manipulate pages via JavaScript; Liu simultaneously reviews Storybook components, builds animations in Remotion Studio, and authors Slidev decks. His preferred deliverable is a single self-contained index.html with JS/CSS — zero deploy. Connectors and Skills package successful workflows as reusable templates for future invocations.

Remote Continuity

Recent updates allow Codex to continue working after screen lock, with mobile monitoring, approval, and takeover — effectively extending the agent's shift beyond the human's presence.