Intel / article

AI Engineer Code Summit: OpenAI on New Ways of Working With Codex

Source: latent.space
Author: Richard MacManus
Date: Published Nov 11, 2026

Context

At a workshop at the AI Engineer Code Summit, two developer-experience engineers from OpenAI, Dominik Kundel and Charlie Guo, presented Codex, the company's coding , as work that can move between several places: a developer's own machine, Codex Cloud, and a dot, OpenAI's always-on personal agent. Speaking with Richard MacManus after the session, as recorded in his report for Latent Space, Kundel called the cloud "just a background", "a different executor of where things are happening", and said a local Codex session can be asked to start a task in Codex Cloud, so "they're all sort of merging together."

The cloud part has a rocky history. Codex launched in May 2025 as a cloud agent that could work on many tasks in parallel, and Kundel admitted it took a while for the models and the surrounding software to make that "a really engaging experience." OpenAI relaunched Codex Cloud on September 30, 2026 with reusable environments, the setup an agent needs before it can work, such as the repository and its dependencies. The workshop writeup puts it as publishing an environment once and running multiple tasks "in isolated instances of that setup." Moving a task that started locally into the cloud is still awkward. MacManus told Kundel he finds it confusing, and Kundel replied, "We definitely want to have it be easier," adding that part of the difficulty is managing the differences in between the two. His own workaround is to ask Codex to push a branch and start a new cloud thread.

The other two topics were ways of steering that work. Guo said most people who code by voice treat it as "just a different medium" on top of their normal work, while he uses it as a "control plane" for managing threads, checking on progress and starting new ones. Kundel said he talks to his dot, which can reach his dev boxes, laptop and cloud environments, on his way to work, so pull requests are waiting for review when he arrives. Guo was more guarded about dots, calling them "a little bit awkward" for Codex power users, though useful as a safety net that flags missed things and as a place to hand off work, which he does "a ton in Slack."

These are two OpenAI engineers describing their own habits in a workshop and, for Kundel, a follow-up conversation with the reporter. The report gives no usage or reliability figures, and a smoother handoff between local and cloud work is something OpenAI says it wants rather than something shown working. Kundel's closing advice was not to get attached to any current way of working, because "things are constantly changing."

Key takeaways · AI-distilled

OpenAI's Dominik Kundel said Codex's cloud mode long felt awkward because models and harnesses took time to catch up. Codex Cloud v2, announced at DevDay, lets you publish an environment once and run many tasks in isolated copies of it, shareable across a company.
Moving a local Codex task to the cloud is still clumsy, Kundel conceded. His workaround is asking local Codex to push a branch and start a new cloud thread, and he describes the cloud as just another executor that local Codex can dispatch work to.
Charlie Guo uses voice as a control plane rather than for dictation: he manages and checks in on threads, pulls data from other apps and starts new threads by speaking, instead of narrating a single lane of work.
Guo said dots, OpenAI's personal agents, feel awkward to Codex power users. He finds them useful in three ways: as a safety net that flags what he missed, as a delegation sink he uses heavily from Slack, and as a presence across everywhere he works.
Kundel starts many tasks from his dot instead of Codex because it already has context from his dev boxes, laptop and cloud environments. He talks to it while driving to work to start PRs that he reviews when he arrives.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.

Read the source latent.space