ChatGPT + Codex Super App

Source & Credit

Summary and action notes from the Agent Native conversation between Riley Brown (host, marketer/founder lens) and Ross Mike (developer / Codex practitioner). Product surface: the merged ChatGPT + Codex desktop “super app,” GPT‑5.6 Sol / Terra / Luna, Sites, and the in-app browser.

Original post: x.com/i/status/2076364713445724178

Companion page for finished-deliverable Work demos: ChatGPT Work summary & action guide. This page focuses on the platform + power-user workflows (Codex tab, computer use, loops, multi-thread, Sites).

What OpenAI Shipped (Big Picture)

1. One super app: ChatGPT and Codex live in the same product — chat, agentic coding, Work, Sites, browser.
2. Three GPT‑5.6 models: Sol (most capable / long-running), Terra (balanced everyday), Luna (fast/cheap).
3. Sites: hosted vibe-coding surface (shareable sites; was teams-only, now broader) — Codex as a full build-and-ship loop.
4. In-app browser upgrades: multi-tab; effort consolidating here as Atlas is sunset / folded in.
5. Work tab: mainstream “get work done” on-ramp for hundreds of millions who still treat ChatGPT as Q&A.

Work Tab vs Codex Tab

Strategic read: Work is how OpenAI gets ~hundreds of millions toward Codex-class capabilities. If you already live in agentic coding, don’t force a tab switch for optics. Details on Work deliverables still live on part 1.

Models: Sol, Terra, Luna (and Fable Reality Check)

Token efficiency > sticker price

• OpenAI’s edge called out: cheaper than many Anthropic options and disciplined tool calling.
• First model test Ross uses: how many tool calls does it make? Agent “psychosis” (15–20 useless tools) kills agentic work even if IQ looks high.
• Gemini: smart on benchmarks, often spirals in agent loops (per this conversation).
• 5.5 was already good at efficient tool use; 5.6 steps that up and finishes multi-phase plans with less babysitting.

Cost-aware routing (don’t drink all the max-model Kool-Aid)

• Fable-class: rare, hard, multi-month-in-one-night builds (sandbox infra, game servers, big ports).
• Sol 5.6: most serious app work at better $/outcome; same ballpark cost story as 5.5 for a smarter, hungrier model.
• Landing-page color change ≠ Fable max. Throttle effort or use a smaller model.
• Subsidies end someday — train the habit of matching model to task now.

Riley’s “Lovable / Replit clone” benchmark: Fable one-shot with a short prompt; 5.6 needed a longer, more specified prompt. Fair comparison is cost + guidance vs raw mythos intelligence — not “who wins Twitter screenshots.”

Power-User Stack: Four Levers

Ross’s practical list for Codex + Sol (and Riley’s marketing extensions of the same pattern).

1. Computer use (biggest workflow unlock)

• Fill forms, drive UIs, QA the app you’re building — often without a custom skill.
• QA loop: build feature → success criteria = computer-use E2E pass → if fail, fix → repeat → only then return.
• Ask for a test plan from app context, then execute and fix failures.
• Background computer use: Codex can run computer/browser work while you keep working (vs Claude computer use taking over the desktop with a blocking overlay).
• Multiple browser instances + virtual terminal in parallel is normal now.
• Ross went so far as a mini PC for remote computer-use agents — think fleet, not one mouse.

Homework (from the episode): spend at least 30 minutes spamming computer use — random tasks + full app QA. You have to feel it.

2. Record & replay → skills (non-dev / marketing gold)

• Tell the agent to record your screen (up to ~30 minutes), then convert the recording into a skill it can re-run.
• Future of knowledge-work computer use: not perfect SOP prompts — show once, replay forever.
• Prompt pattern: “Record my screen and turn what I do into a skill.”
• Ideal for repetitive 10-minute admin/marketing flows you’d never bother scripting by hand.

3. Loops (dev + marketing)

Simple definition used on the show:

1. Well-defined prompt (what to produce)
2. Feedback engine (how to judge quality)
3. Success criteria (when to stop)

Keep looping until criteria hit. The agent doesn’t invent success for you — you must define it. See also loops, not prompts and stop babysitting agents.

Why coding still trains better than pure design: pass/fail is clearer. Marketing loops fix ambiguity by anchoring to high-quality examples. More Content OS patterns: AI content marketer.

4. Tools / building-block stack

• Speed > owning every primitive. Agents prefer composable blocks over raw AWS from scratch.
• Ask Codex what tools it prefers — it will tell you what it can use well.
• Examples from the episode: Convex (code-first backend), Daytona sandboxes, Vercel AI SDK / Eve agent framework, domain MCPs (e.g. Blender).
• Agent-friendly frameworks use descriptive folders (agent.ts, skills/, tools/, sandbox/, channels/) so models navigate by filesystem semantics.
• Economy shift: plugins, markdown, MCPs designed so one prompt can assemble Legos — humans still maintain the blocks.

5. Multiple threads (controversial, high leverage)

• Compartmentalize non-overlapping work; stop serializing everything into one fat context.
• Fresh thread = full capability; a thread already at ~120k tokens is a degraded agent.
• Codex can spawn many threads from one: “spin a thread per feature, computer-use test each, report results” → leave and come back.
• Meta-prompt: summarize all of today’s threads into a one-pager with links back into each session (builder or marketer OS).

Sites + In-App Browser = Vibe Coding Platform

• Sites: create/host shareable websites inside the product (Hello World → real apps). Iterate in-product; promote to Vercel later if you want.
• From Codex: new chat + @sites (or Sites UI) — model gets the Sites plugin surface automatically for many website asks.
• Browser panel: full in-app browser; multi-tab support closes a major previous gap.
• Atlas as a separate browser effort is being sunset / redirected into making this embedded browser excellent — agents open pages here instead of a detached Chrome side quest.

Related mobile/agentic build notes: Mobile apps with Codex.

Super App, Agent-Native Mini-Apps, AI OS

• Super app (Codex/ChatGPT shell): chat + tools + browser + sites + computer use ≈ early AI operating system.
• Agent-native / mini apps: UIs you rarely open alone — the agent surfaces them (e.g. email drafts to approve/send). Human + agent co-control.
• Trajectory: buy a computer with this preloaded; talk → something runs or an interface appears → next chat. Early days, already real.

Action Checklist

Mindset Closes from the Episode

• Imagination is the bottleneck once Sol-class models + computer use + schedules exist.
• Most products are markdown files — don’t spend your ambition building another notes UI; build something useful and hard.
• Education gap > tool gap. Tools move faster than adoption; terrible short-form advice is common. Teaching real workflows (ops/marketing loops, plugins, OS design) is the opportunity.
• Riley’s focus: agentic workflows for non-coding work at scale. Ross’s focus: dream backlog + multi-thread building, while leveling design/writing taste.

Related Guides

Summary adapted from the Agent Native conversation with Riley Brown and Ross Mike on the ChatGPT + Codex super app, GPT‑5.6 models, computer use, loops, Sites, and agent-native workflows. View the original post on X → · Companion: ChatGPT Work part 1.