Codex cooking field note

Cooking with Codex after AI Engineer World's Fair.

The useful conference recap is not a recap. It is a runnable way to pick the right Codex surface, build a demo in the browser, package the learning as a skill, and leave behind a report artifact that another project can reuse.

Opinions Cooking with Codex after AI Engineer World's Fair.
Join newsletter

I refreshed the current AI Engineer World's Fair sources on June 30, 2026: AI Engineer World's Fair schedule v4844, 568 sessions, 549 speakers, and official tracks covering Computer Use, Voice & Realtime AI, Context Engineering, Local AI, Software Factories, Claws & Personal Agents, Vision & OCR, and Harness Engineering. The session list includes Cooking with Codex with Charlie Guo and Gabriel Chua. That makes the right question very concrete: what should a builder actually cook?

Illustrated Codex workbench with browser preview, terminal receipts, goal dashboard, and plugin cards.
Generated visual receipt for this note. The voice model is gpt-realtime-2; the raster image workflow belongs to image generation, not Realtime.
The operating rule:

A Codex demo is done only when the goal, source receipts, chosen tool plane, visible state change, verification command, and reusable skill/reference packet all exist in the same artifact chain.

Start from the method, not the plugin list

I am using "the Matias Duarte method" here as a design lens, not as a sourced quote: begin with the human state transition. What does the person know, what can they safely do next, what feedback appears, and what remains reversible? If the demo cannot answer those questions in the first minute, the model choice is premature.

The "Charlie Guo method" is the Codex cooking lens I would use from the World's Fair prompt: do not ask one giant agent to be impressive. Give Codex a clear goal, a current appshot or browser surface, the narrowest useful tool plane, and a proof step it must run before the story counts. The public schedule confirms the Codex cooking session; this note turns that pressure into the actual operating loop.

The review board I would use

This is an inferred operator review board, not a set of quotations or endorsements. I would use these public-source lenses to make a Codex demo harder to fool and easier to reuse.

Charlie Guo

Make the agent do the thing.

Source lens: the World's Fair schedule pairs Cooking with Codex with Charlie Guo and Gabriel Chua, and also lists Charlie's Voice Agents Can Just Do Things. The review question is whether the agent performed a visible action, picked the right surface, and left proof.

AIE sessions JSON
Garry Tan

Stress-test the idea before code.

Source lens: YC describes GStack as skills for office hours, design, code review, QA, and browser testing across Claude Code, Codex, or Cursor. The demo should start with adversarial office hours, then move through design, code review, and QA in one loop.

YC GStack walkthrough
Sam Altman

Find the users who love it.

Source lens: the Startup Playbook keeps product love, clear users, focus, and execution quality ahead of demo theatrics. The Codex cooking pass should identify a narrow power user and prove the workflow saves that person real time.

Startup Playbook
Dario Amodei / Anthropic

Raise capability and safeguards together.

Source lens: Anthropic's Responsible Scaling Policy treats more capable models as requiring stronger risk reports, security controls, and deployment safeguards. A Codex demo that gains tool authority should add evals, boundaries, rollback, and misuse checks at the same time.

Responsible Scaling Policy

Computer Use vs Chrome vs Browser

The most useful Codex decision is often not "which model?" It is "which surface should Codex operate through?" The wrong surface creates fake difficulty. The right one makes the demo shorter and more inspectable.

Surface Use it when Receipt to demand
Codex native Browser Localhost, file-backed previews, public pages, or a web demo that does not need a signed-in Chrome profile. Screenshot, rendered DOM check, read-only JS probe, and route/search smoke result.
Chrome extension Logged-in websites, browser profile state, Chrome extensions, or tools where the real user session matters. Named profile context, visible signed-in state, and a non-destructive action receipt.
Computer Use Desktop apps, Mac apps, non-browser UI, visual-only workflows, or cases where no structured integration exists. Original-detail screenshot loop, click target evidence, and a before/after state change.
Dedicated app/plugin GitHub, Gmail, Drive, Calendar, OpenAI docs, Convex, Mac apps, iOS app scaffolding, or any structured API path. Tool response, file diff, exported artifact, or typed API evidence instead of fragile UI clicking.
Realtime and media Use gpt-realtime-2 for low-latency voice agents; use gpt-image-2 and vision for visual artifacts. Voice session trace, interruption behavior, generated image asset, or vision analysis with source image.

The top five frontier cooks from 2026 Q2+

The interesting pattern from the current docs and conference data is not one tool. It is the convergence of execution surfaces, packaging, live UI proof, and stateful backends. These are the five I would copy first.

  1. Codex as a multi-surface execution plane. The desktop app, cloud tasks, appshots, subagents, worktrees, Browser, and connected hosts let one goal move between local code, rendered UI, and remote execution without collapsing into chat history.
  2. The Browser, Chrome, and Computer Use split. Browser is for public/local web surfaces, Chrome is for signed-in web state, and Computer Use is for desktop UI. This distinction is now a core product skill.
  3. Agent Skills as the reusable packaging layer. The agentskills.io layout gives every serious workflow a SKILL.md, optional references/, scripts/, and assets/. That is how a one-off demo becomes a sibling-project capability.
  4. Reactive agent backends. Convex's Agent component describes message history, vector search, long-running workflows, and real-time database reactivity in one app surface. That is the shape of a useful progress dashboard, not just a chat widget.
  5. Separate realtime voice from visual generation. Realtime is the live voice/tool loop. Image generation and vision are the visual artifact loop. Mixing them in product copy creates implementation debt.

The Define Goal stack

The "Define Goal" skill should not be a slogan. It should create a series of goals, each with an owner thread, source receipts, status, artifact, and audit question. In Codex App terms, that is the bridge between a good prompt and parallel work that can be reviewed.

Powered by Agent Skills

Goal dashboard for cooking with Codex

The standalone report artifact lives at memory-bank/artifacts/codex-worlds-fair-cooking-2026-06-29/progress-dashboard.html.

Goal 01 Source map

Pull official docs, AIE schedule data, Agent Skills spec, and Convex receipts.

Done: source links and schedule v4844 captured.
Goal 02 Tool chooser

Decide Browser, Chrome, Computer Use, plugin, Realtime, or image/vision before acting.

Done: table shipped in this route.
Goal 03 Browser demo

Build the public note and dashboard so a user can inspect the claim in the browser.

In progress until live-domain verification passes.
Goal 04 Remote setup

Document Codex cloud and connected-host setup as separate execution paths.

Done: setup checklist below.
Goal 05 Skill promotion

Move reusable choices into skills/codex-cooking-workflow/references/.

Done: sibling projects can reuse the packet.
Goal 06 Goal audit

Every spawned thread needs a blocker, artifact, verification command, and next repair step.

Queued: use this as the follow-up audit rubric.

How to set up remote Codex execution

Codex has two different remote stories. Codex cloud runs tasks in configured cloud environments connected to a repository. Remote connections let a trusted host provide files, shell, tools, credentials, plugins, Browser, and Computer Use to another Codex surface. Treat them as different products.

Cloud task

Use Codex cloud for repo-contained work

Configure the environment at ChatGPT Codex settings, connect GitHub, define setup commands, secrets, caches, and domain allowlists, then launch tasks with logs, diffs, and reviewable outputs.

Connected host

Use remote connections for a real machine

Keep the host awake, install and authenticate Codex App, pair mobile or desktop, and rely on SSH host entries for dev boxes. Do not expose the app-server publicly.

Chopshopr product

Make setup deterministic

A remote Chopshopr host needs bun install, bun run build, MCP registrations, OpenAI docs access, deploy credentials, and local model health checks documented as receipts.

Customer path

Show status before delegation

The frontend dashboard should show source freshness, environment readiness, current goal, spawned thread, artifact path, and last verification command before the agent starts cooking.

The app-server protocol layer

If remote connections are the product experience, Codex app-server is the lower-level protocol surface for building a rich Codex client inside your own product. It is not the CI automation path; OpenAI's docs point automation jobs toward the Codex SDK instead. App-server is for authentication, conversation history, approvals, streamed agent events, and thread/turn control.

Chopshopr rule:

Treat app-server as a private control plane. Put product auth, tenancy, audit logs, and user-visible status in Chopshopr. Keep raw app-server listeners local, on a Unix socket, behind stdio, behind SSH port forwarding, or behind explicit WebSocket auth and TLS. Never expose an unauthenticated non-loopback listener as the product API.

Protocol part What matters Product receipt
Transport JSON-RPC 2.0 style messages over JSONL stdio, WebSocket, Unix socket WebSocket, or off. Chosen listener, local binding, token/TLS story, and health probe result when WebSocket is used.
Handshake Every connection starts with initialize, then the initialized notification. Client name, version, capabilities, and any notification opt-out recorded in logs.
Work model Threads hold conversations; turns hold user requests; items stream messages, tool calls, files, and command work. Thread id, turn id, active status, artifact path, and final turn/completed state.
Events Clients keep reading notifications such as item start/completion, agent-message deltas, tool progress, interrupts, and turn completion. Progress dashboard that shows live event state instead of a fake spinner.
Backpressure WebSocket mode uses bounded queues and can reject overloaded ingress with a retry-later JSON-RPC error. Exponential backoff with jitter, idempotent client actions, and visible retry state.
Versioning Generate TypeScript or JSON Schema bundles from the exact Codex version you run. Committed schema artifact, compatibility test, and upgrade note before changing clients.

Best practices for everybody building on AI

  • Choose the tool plane before the model plane. Bad surface selection creates most demo friction.
  • Make every appshot actionable: current state, intended action, safety boundary, and expected receipt.
  • Prefer structured plugins and connectors before UI automation when a trustworthy API exists.
  • Use app-server for rich product clients only after you can explain transport, auth, lifecycle, event streaming, backpressure, and schema versioning.
  • Use Computer Use for real desktop surfaces, and keep screenshot detail original for click accuracy.
  • Keep voice, image, and browser demos separate until each one has its own verification path.
  • Promote every repeated lesson into a skill reference with scripts or assets when possible.
  • Let progress dashboards show the audit trail, not just a pretty completion percentage.

Questions for the end reasoner

The move is not from one better prompt to one larger agent. It is from prompting an agent to designing the loop the agent can keep running, auditing, and improving.

Question 01

Have I tried asking Codex to do it?

Give Codex the artifact, the success test, the allowed tools, and the permission boundary. If it still stalls, the blocker is probably not model intelligence.

Question 02

What is stopping Codex from getting to the answer?

Classify the failure: missing source, hidden state, unavailable environment, unsafe action, absent test, or ambiguous taste call. Each class gets a different repair.

Question 03

Why does this process need a human at all?

If the human is only applying a repeatable preference, encode the rubric. Keep the human where taste, accountability, access, or irreversible action is actually the job.

Question 04

How do I move from prompting agents to building loops?

Convert the prompt into a DMAIC loop: define the goal, measure the current state, analyze the blocker, improve the artifact, and control recurrence with a skill, test, or dashboard.

Question 05

How do I scale from one agent to many?

Split by artifact boundary, not vibes: source scout, implementer, reviewer, browser verifier, deploy verifier, and goal auditor. Merge only through receipts.

Question 06

How do I build the machine that builds the machine?

Promote the winning loop into skills, scripts, app-server events, progress data, search entries, smoke tests, and reports that sibling projects can run without the original author in the room.

The related Chopshopr threads are already pointing at the same operating system: Skillable is scalable for reusable skills, Nothing left inside the tool call for action preflight, Applause instead of adoption for explicit state and authority, Arthasida Frontier Console for control-loop review, and Reasoning Quality Lab for turning human methods into package gates.

The demo to cook first

The smallest useful browser demo is the one this route now carries: a source-backed blog post, generated visual, tool chooser, goal dashboard, report artifact, search entry, smoke-test needles, and a reusable skill reference. It is not the flashiest possible Codex demo. It is the one another project can absorb tomorrow.

Browser demo receipt bun run build open https://chopshopr.ai/blog/cooking-with-codex-worlds-fair/?v=20260629 open https://chopshopr.ai/search/?q=codex%20cooking%20dashboard

The local build proves route, search, smoke, and tests. The live open proves the deploy. The search query proves the learning is discoverable instead of buried.