At the AI Engineer Code Summit, workshops from OpenAI, Anthropic and Google DeepMind opened the event. I attended OpenAI’s session, titled "New Ways of Working with Codex," which was led by Developer Experience engineers Dominik Kundel and Charlie Guo. The session focused on recent Codex developments and how three areas — cloud execution, voice controls, and personal agents called dots — fit into developer workflows.
What has changed so far
Charlie Guo opened by listing features shipped to Codex since the AI Engineer World’s Fair in June 2026 and noted the team has been releasing updates at a very rapid pace. The presenters also acknowledged that not every change has been seamless, and the migration to cloud-based workflows has been a particularly rough spot for many users.
The awkward move to the cloud and Codex Cloud v2
Codex launched in May 2025 as "a cloud-based software engineering agent that can work on many tasks in parallel." Kundel admitted it took time for models and execution harnesses to align into a consistently engaging experience — meaning the cloud component of Codex often felt awkward for users.
That began to shift after DevDay in late September, when Codex Cloud version 2 was announced. The update aims to make cloud usage a practical option: the official description says you can "publish an environment once and run multiple tasks in isolated instances of that setup." Kundel explained those cloud environments can be shared across a company or used individually, and you start by selecting the GitHub repositories the environment should access.
After the workshop I told Kundel I still find some aspects of Codex’s cloud confusing, such as moving a task started locally into the cloud. He acknowledged the issue and said OpenAI wants to make that process easier; some fixes depend on managing differences in context. He demonstrated reopening an existing cloud task in the web UI and showed that its Appshot (a feature for saving an app experience into ChatGPT) remained intact. For a task begun locally, he described pushing a branch and starting a new cloud thread.
Kundel said Codex Cloud is increasingly acting as a background executor — an alternate place where things run — and that local and cloud Codex instances are "merging together."
Voice as a user interface
One demo showed live edits to a fictional record label site, Blossom Music; Kundel said he sometimes uses voice dictation when iterating because it lets him iterate quickly when he isn’t sure of the final direction.
Guo later explained voice is also effective for delegating work to different agents. He observed most people treat voice as simply another medium layered on top of their usual work, but he personally uses voice as a control plane: to manage threads, check on tasks, pull data from other apps, and create new threads in the application. In short, he uses voice not only to input code but to orchestrate the development session.
Kundel added that beyond dictation he frequently talks to his dot. Because his dot has access to his dev boxes, laptop and cloud environments, it holds context about what he works on. He said he often talks to it on his way to work to process items and start PRs so that they’re ready for review when he arrives at the office or returns home — illustrating how powerful voice can be.
Dots and their role
Personal agents, or dots, were discussed briefly. Guo conceded that dots can feel awkward for seasoned Codex users because they represent a new paradigm and excel at different tasks than the main Codex.
He described three practical roles where dots are useful: 1) as a safety net that can flag missed issues, 2) as a "delegation sink" — something he uses heavily in Slack for delegating tasks, and 3) as an ubiquitous presence providing context everywhere you work. The goal, he said, is to make dots genuinely available across environments where developers operate.
Kundel said he often starts tasks from his dot rather than prompting Codex directly because the dot already has a lot of context and can push PRs forward on his behalf.
Conclusion
Kundel closed the session with practical tips for working with Codex — for example, lowering the agent’s effort level and asking agents for help — and reminded attendees not to become too attached to current workflows because "things are constantly changing."
Overall, OpenAI’s presentation highlighted a fast-evolving Codex ecosystem: cloud execution, voice controls and personal agents offer new ways to work, but usability and context handoff remain areas under active development.
