Codex + Agent OS — automate anything.

Codex Thursday — what just shipped

OpenAI dropped four updates yesterday. Each one nudges Codex from "helpful coding assistant" closer to "autonomous workforce sitting on your Mac." Stack them together and the result is something new.

The headline is Goal Mode. Always-on context, persistent execution, hands-off operation for hours or days. That's a different category of tool than "chat with a coding model."

Why standalone Codex still leaves money on the table

OpenAI built an incredible agent. But the way it ships, you still have to open a separate app to use it. Five tabs across your screen. No memory of what your other agents (Hermes, Antigravity, Claude, OpenClaw) are doing. No shared workspace. No preview of what it built without context-switching to a browser or file manager.

Goal Mode is the killer feature — but if your goals live inside one app, your sessions inside another, your workspace files inside Finder, and your other agents in five more tabs, you're not running a system. You're juggling.

Wire Codex into Agent OS

I built Codex into Agent OS as a first-class sibling — sitting in the sidebar next to Hermes, Antigravity, Claude, Gemini, OpenClaw, and Free Claude Code. Same Mission Control. Same memory layer. Same preview window.

Under the hood, it's codex exec --json for streaming chat and codex exec --full-auto --json for Goal Mode, against your default Codex profile. Same auth as your terminal. Same model. Same sessions. Just wrapped in a workspace your eyes can actually use.

The integration gives you four tabs:

Chat — single-shot, multi-turn memory

The Chat tab is where most quick Codex work happens. Behind the scenes it's codex exec --json streaming each turn back to Mission Control, with prior conversation packed into every new prompt so Codex doesn't lose the thread.

Chat tab. CODEX · DIRECT · codex exec --json. Multi-turn memory, same auth + model as your terminal codex, Esc to abort an in-flight call.

For long-running work, switch to Goal Mode. For one-shot questions or fast iteration on a small task, this is where you live.

Goal Mode — the hands-off engine

This is the centerpiece. The thing OpenAI shipped that nobody else has shipped. The reason this whole guide exists.

You hand Codex a long-horizon objective. Something like "Build a single-file HTML landing page for an AI cookbook subscription, generate the imagery, write the copy, optimise the meta, and save it ready to ship." You click Launch. You close your laptop.

Codex runs codex exec --full-auto --json in a dedicated working directory under ~/codex-scratch/<id>/ so artefacts from different goals never collide. It works. It iterates. It checks. It writes files. It keeps going.

Goal Mode. Drop a long-horizon objective on the left, watch live log on the right. Each goal runs codex exec --full-auto --json in its own scratch dir.

Goals persist to ~/.agentic-os/codex-goals.json so they survive reloads and dev-server restarts. Status surfaces as queued, running, completed, failed, or stopped. You can have multiple goals running in parallel — each isolated in its own scratch dir, each tracked independently.

The Goal Engine — patient, autonomous, mythological. Each waystation is a checkpoint, each artefact a finished step.

Sessions — your past 63 runs, surfaced

Codex stores every session on disk. The CLI does this automatically. But the CLI doesn't give you a good way to browse them — and you certainly can't see them next to your other agents.

Agent OS reads ~/.codex/session_index.jsonl and lists every past session newest-first. Thread name. Session ID (mono). Relative timestamp.

Past sessions, newest first. Set up Hermes MCP · Create cat image · Use browser · Find AI automation SEO keywords · Create AI automation blog · Open calculator and type something · Create Codex business presentation · Summarize recent Gmail requests · and 55 more.

Why this matters: every past session is a template. Every successful workflow is a record. If you nailed "find AI automation SEO keywords" three weeks ago, you can re-open it, see what worked, fork it for a new niche. Compounding institutional memory — but only because you can actually see it.

Workspace — preview anything Codex builds

This is the tab that closes the loop. You launch a Goal in Tab II. Codex spends an hour building a landing page, generating images, writing copy. Where does the finished work go?

Anywhere standalone Codex puts it, you'd have to dig through Finder, open files manually, switch between apps. With Agent OS: it appears in the Workspace tab.

Workspace. Projects from ~/codex-scratch/ on the left. Pick one, see its files. HTML renders in an iframe. Images preview inline. Videos play. PDFs embed.

Same pattern as the Antigravity workspace. Same preview/source toggle. Same path-based serving so the iframe just works. The result: you launch a goal, walk away, come back, click the project, see the thing. No context-switching. No file-hunting.

The Workspace as a gallery. Every artefact your agents produced, in alcoves you can survey at a glance.

Where Codex sits in the agent stack

Codex isn't alone in Agent OS. It joins a roster of agents, each with its own strengths:

• Claude — reasoning, long-form planning, anything that needs depth.
• OpenClaw — local-first always-on AI assistant, lives across WhatsApp, Slack, Discord.
• Hermes — Nous Research agent, tool calls, kanban, plugins, skills.
• Gemini — long context, multi-million-token jobs, fast multimodal.
• Antigravity — see The Goldie Gravity Grid → — Google's agent-first deployment platform.
• Codex — autonomous Goal Mode, locked-Mac control, OpenAI's hands-off engine.
• Free Claude Code — the OSS sovereign-stack route. See Free Claude Code + The Sovereign Stack →

Each pulls from the same Obsidian memory layer. Each has its own workspace, its own session history, its own goal queue. You stop juggling tools. You start orchestrating fleets.

3 things I actually have Codex doing right now

Straight from my Sessions tab — these aren't hypotheticals. They're live goals running on my machine this week.

• Set up Hermes MCP — Goal: wire a new MCP server between Hermes and Obsidian. Codex bootstrapped the config, ran the install commands, tested the connection, and reported back. Two hours, no babysitting.
• Find AI automation SEO keywords — Goal: pull keyword volumes for "AI automation" cluster, deduplicate, score by difficulty, output a CSV. Walked away, came back to 142 keywords ranked and ready to draft against.
• Summarize recent Gmail requests — Goal: skim last 7 days of inbox, surface anything requiring a response, draft replies for the urgent ones. Standing weekly automation. Saves an hour every Monday.

None of these were "write some code." All of them were achieve this outcome. That's the Goal Mode unlock.

Three beliefs holding you back

The Codex updates land differently depending on which beliefs you hold. Three I see most often:

30-day roadmap

Don't try to automate everything at once. Build the muscle. Start small, expand as you trust it.

The Goldie Goal Engine™ — recap

Codex Thursday gave us four launches, and the biggest one is Goal Mode. Locked-Mac control + Appshots + Goal Mode + Annotation together make Codex a different category of tool — not "chat with code" but "autonomous worker that finishes outcomes."

But OpenAI shipped the engine, not the vehicle. Standalone, you're still juggling tabs and losing sessions. Wired into Agent OS — four tabs, the past 63 sessions, the workspace preview, the sidebar next to every other agent you run — Codex becomes the Goal Engine: hands-off, persistent, compounding.

One agent in. Four tabs of leverage out.