New Ways of Working with Codex

Dominik Kundel, OpenAI, Charlie Guo, OpenAI58:03 · Nov 2026 · 6,955 views
Thumbnail for New Ways of Working with Codex Watch on YouTube
TL;DR
  1. 1

    Codex now shares one harness across ChatGPT, the CLI, Slack, Dots, the Agents API, local devices, and cloud environments.

  2. 2

    Plugins, Appshots, computer use, computer history, and ChatGPT Space let agents gather context from code, documents, applications, and team discussions.

  3. 3

    Codex Cloud, ChatGPT Sites, voice control, code review, and parallel agents extend Codex from coding tasks into shipping, review, and coordination work.

Summary

Dominik Kundel and Charlie Guo describe Codex as a shared agent harness that has expanded across ChatGPT, local machines, cloud environments, Slack, Dots, and the Agents API. The talk focuses on reducing the amount of context users must spell out. Plugins connect agents to external systems, Appshots add accessibility-tree data to screenshots, and ChatGPT Space gives people and agents a shared editable workspace. The speakers also show fast UI iteration with Ultra Fast and annotations, long-running work in Codex Cloud, and a ChatGPT Site called Cut Room that sends footage to a remote agent with access to DaVinci Resolve. Later sections cover code review, voice control, Dots, reasoning effort, model selection, and parallel agents. Their advice is practical: use high reasoning as a starting point, lower it when appropriate, and let agents delegate work when tasks can be split.

Key ideas
00:12

Codex has become a shared harness across many work surfaces

Charlie Guo says Codex has evolved rather than disappeared. After Codex merged with ChatGPT work, more people could use its long-running agents. The same harness now powers ChatGPT work, the CLI, ChatGPT, Slack, Dots, and the Agents API. Charlie also describes a return to Codex's original cloud-native shape, with agents working in remote environments while users control them from other devices. The talk is organized around context, iteration, local and cloud work, shipping, delegation, and everyday operating tips.

03:59

Plugins let agents gather context from systems outside the repository

Dominik Kundel argues that coding agents need the goal and surrounding context, not just a precisely written prompt. Repository standards belong in the repository through files such as agents.md, linters, and other tools. Information outside the repository can come through plugins, which bundle MCP servers, skills, and other resources. A Slack connection can be reused across the web, a developer box, and Slack itself. Kundel describes using Slack, Google Drive, and shared spreadsheets so Codex could shepherd a documentation launch, watch for last-minute changes, and confirm that the site was live.

12:57

Appshots and computer use add application context beyond a screenshot

Appshots capture a screenshot together with accessibility-tree information from the page. That gives Codex details outside the visible viewport, such as text, variables, and numbers. Computer use can interact with applications through its own cursor, including an iOS simulator. A Chrome extension can provide logged-in website context on Windows, Linux, and macOS without taking over the whole desktop. On macOS, computer history records application activity and can help Codex learn habits, such as where blog posts are drafted. Users can pause history, clear recent activity, and exclude apps or websites.

17:52

ChatGPT Space gives people and agents a shared place to revise plans

Kundel describes ChatGPT Space as a way to keep plans out of a copy-and-paste loop between a repository, a Google Doc, and a team discussion. Codex can turn a plan into an editable rich-text page. People can tag ChatGPT or a colleague, while agents can update the page and add visualizations or HTML objects. Assistant instructions can tell an agent how to modify a page, and auto-update can keep information current. Kundel compares these page-level instructions to agents.md files for individual pages, especially for documents such as on-call runbooks.

20:57

Fast UI work depends on short feedback cycles and the right reasoning setting

Ultra Fast is presented as useful when a developer is exploring an interface and wants to stay in an interactive creative loop. Kundel warns that it will not always feel faster when compilation, builds, or end-to-end tests dominate the task. He recommends a hot-reloading development server and usually light reasoning for this style of work. In the demo, he gives small instructions to change a page, adds animation, makes a vinyl record spin, and asks for more readable elements. Annotations let users select specific parts of a page and provide targeted feedback, while an annotations API defines what should be selectable.

28:27

Codex Cloud prepares reusable remote environments for long-running work

Kundel says long-running tasks have led people to keep Mac Minis, old laptops, or rented servers running, but maintaining those machines becomes difficult. Codex Cloud creates environments from one or more GitHub repositories. It works out dependencies, API keys, startup commands, and verification commands, then snapshots and updates the environment so later tasks can start quickly. The environment can store personal or shared secrets, control web access, and reach private services through OIDC or Tailscale. Linux environments cannot build an iOS app, but they can retain the mobile repository for reference. Codex CLI can also expose a remote dev box without requiring the desktop app.

37:08

ChatGPT Sites can turn Codex work into full internal applications

Charlie Guo uses Cut Room, an internal ChatGPT Site, to manage launch-video production. A team uploads footage, leaves comments, adds screen recordings or B-roll, and submits a cut for review. A remote Codex agent then accesses DaVinci Resolve and video-editing skills, creates a new cut, uploads it, and notifies the team. Guo says Sites include database service through D1, file and object storage through R2, authentication through Sign in with ChatGPT, plugins, analytics, and secrets. He also describes using plugins with Google Calendar and Linear. Sites can present internal data as interactive experiences rather than static documents.

43:36

Delegation works across review, voice control, Dots, and custom products

Codex's code review tab can find connected GitHub pull requests, show diffs and screenshots, and start a review with a prepared prompt and custom instructions. Guo describes voice mode as a control layer for threads and applications, rather than only dictation for one coding session. Dots can act as a safety net, a place to send delegated tasks from Slack, and an agent reachable through phones or messaging. The beta Agents API brings the Codex harness, sandboxes, and plugins into other products, including applications that need many temporary sessions.

51:56

Reasoning settings and parallel agents should be chosen for the task

Kundel recommends starting with high reasoning instead of automatically choosing the highest setting, since newer models may perform well at lower effort and higher effort can produce diminishing returns or overthinking. An agent can use one model at medium effort and ask other models to handle harder or more tedious subtasks. Peer agents can work on related parts of a project, such as web and mobile applications, while sharing relevant backend changes. Separate chats can also run independently, and the CLI now provides an agents view for switching between sessions.

"You really want to be able to converse with your coding agent the same way that you would do with a colleague or a friend who understands and has all of the context already."Dominik Kundel04:21
Who should watch
  • You are building coding workflows that need context from Slack, documents, design systems, or browser applications instead of one repository alone.
  • You want to run coding or review tasks remotely while keeping local machines free, or you need a repeatable environment for long-running work.
  • You are experimenting with agent delegation, voice control, ChatGPT Sites, or parallel sessions and want concrete workflow examples.