Dominik Kundel and Charlie Guo, from OpenAI’s developer experience team, demonstrate connected Codex workflows on AI Engineer.

Give the agent access to the right context

  • Keep architecture, coding conventions and build commands in the repository, alongside AGENTS.md and automated checks.
  • Plugins bundle skills, resources and MCP tools, which connect an agent to outside services such as Slack and Google Drive.
  • Their design-system plugin adds a component gallery and inserts selected components directly into the prompt.
  • Appshots combine a screenshot with accessibility text, including information outside the visible window.
  • Computer History learns work habits from activity on macOS; the presenters describe pause, deletion and app-exclusion controls.
  • ChatGPT Space puts a shared plan, team comments and instructions for updating the document in one editable page.

Shorten the feedback loop

  • Use faster generation for small changes you can inspect immediately, with a hot-reloading development server.
  • Long builds and test runs can absorb the benefit of faster model responses.
  • Browser annotations attach feedback to specific elements; a custom annotation API can select whole components or game regions.
  • Kundel iterates on a record-player interface, adjusting its layout, animation and scratch interaction through successive requests.

Put work on the machine it needs

  • Codex Cloud prepares environments from multiple repositories and identifies dependencies, setup commands and missing access.
  • The demo keeps an iOS repository available for reference while acknowledging that the hosted Linux machine cannot build its app.
  • The presenters describe Tailscale for private network access and OpenID Connect for scoped cloud credentials.
  • codex remote-control start makes an existing development host accessible through Remote without installing the desktop app there.

Ship, review and delegate

  • Their internal Cut Room app sends footage and comments to a remote agent using DaVinci Resolve, then returns an edit for human review.
  • ChatGPT Sites combines hosting with database storage, file storage and sign-in; connected plugins add application data.
  • The Code Review view opens pull requests and launches a separate review conversation with adjustable instructions.
  • Voice can start, inspect and steer multiple threads; the live demo asks which “Matt” the speaker means in Slack.
  • Dots serve as a proactive safety net and a place to forward tasks; the Agents API exposes the underlying agent system to applications.

Spend reasoning effort deliberately

  • Re-test lower reasoning settings when models improve; reserve extra effort for tasks that need it.
  • A main agent can ask stronger models for advice and delegate bounded work to smaller models, but routing can add cost and time.
  • Pair agents with complementary responsibilities when web, backend and mobile decisions need coordination.
  • Independent child threads provide another way to run work beyond a single parent-managed subtask.

“Don’t get too attached to how you’re using things.” — Dominik Kundel, 56:55

Caveat: this is an OpenAI product workshop. Availability varies by platform, plan and feature; internal workflow examples are the presenters’ reports.

Workshop resources and examples