Codex Wants to Make Cloud Coding Feel Less Like Cloud Coding

Codex is trying to make the cloud disappear.

At the AI Engineer Code Summit, OpenAI presented new ways to move coding tasks between local machines and cloud environments, steer work with voice, and keep agents running after a developer closes the laptop. The company also used the wider Dev Day roadshow to position its always-on Dots assistant as a proactive enterprise tool, creating a neat collision between developer convenience, compute costs, and privacy.

The summit kicked off with workshops from OpenAI, Anthropic, and Google DeepMind. The OpenAI session, titled “New Ways of Working with Codex,” was hosted by Developer Experience engineers Dominik Kundel and Charlie Guo, who focused on the awkward parts of modern coding agents: what happens between sessions, where tasks run, and how much context survives the trip.

The past year has brought constant product upgrades across the frontier labs, but the shift to cloud remains confusing for users of Codex, Claude, and Antigravity. Personal agents such as OpenAI’s dots and Anthropic’s Claude Duet can assist between sessions, while voice mode gives developers another way to steer coding tasks. OpenAI’s workshop addressed all three problems.

Codex Cloud Wants To Become Background Infrastructure

Guo opened by listing features shipped in Codex since the AI Engineer World’s Fair in June 2026. He said, “I had to cull this list in order to make it fit on the screen because there were too many things to put in here,” adding, “I think the team has been shipping at an unbelievable rate.”

Codex launched in May 2025 as “a cloud-based software engineering agent that can work on many tasks in parallel.” Kundel acknowledged that “it took a while for models and harnesses to equalize into something where we can build a really engaging experience.” That is a polished way of saying the pieces needed time to stop fighting each other.

After DevDay in late-September, OpenAI announced a revamped version 2 of Codex Cloud. Its promise was simple: you can “finally close your laptop now and your agents will keep working.” The new system lets users “publish an environment once and run multiple tasks in isolated instances of that setup,” with environments that “can be shared within a company if you’re working in a company, or you can just create your own.”

Users begin by selecting the GitHub repositories Codex can access. The lingering problem is movement: a developer may start locally, then decide the task belongs in the cloud, without wanting to restart the work or reconstruct its context from scratch.

Kundel said, “We definitely want to have it be easier,” but added that some of the challenge depends on “managing the context differences.” He described asking Codex to hand a local task to the cloud, then reopening an existing cloud task on the web, where its Appshot was still available. Appshot is “a new feature enabling you to save an app experience to ChatGPT.”

In another flow, Kundel asked Codex to push a branch and start a new cloud thread for work that began locally. He described Codex Cloud as “just a background, like a different executor of where things are happening,” and said local Codex can spin up a task on cloud Codex so “they’re all sort of merging together.” The goal is clear even if the interface still needs translation: local and cloud agents should feel like parts of one workspace.

Voice, Dots, And The Cost Of Always-On AI

Kundel demonstrated a new website for the fictional record company Blossom Music and made live edits during the workshop. He said he sometimes uses voice dictation while iterating, allowing developers to describe changes instead of typing every instruction, and argued that this helps users reach an intended result when they are unsure what they want.

That same shift toward continuous assistance sits behind Dots, which OpenAI took to London during its Dev Day roadshow after announcing the product at its annual developer event in San Francisco a month earlier. Tools such as Codex wait for instructions; Dots are designed to learn a user’s goals, watch inboxes and work tools, and begin tasks before the user asks.

Dots are not available in Europe yet, though OpenAI is working with European regulators to change that. The company is rolling them out to enterprises and its most expensive Pro subscribers in the U.S., while Romain Huet said OpenAI wanted to “nail the enterprise and pro use case first.”

The business case comes with an unpleasant invoice. Enterprises spent $177 billion on AI in the year to September 2026, up from $33 billion a year earlier, while the number of companies paying for AI rose only modestly; existing customers supplied most of the growth. On OpenRouter, agents now generate four out of five tokens, and Anthropic, OpenAI, and Google account for more than two-thirds of enterprise AI deployment spending.

Always-on agents also consume serious compute. Musk said Grok Bot’s running costs were “unsustainably high relative to revenue,” while Adam Fry described Dots as “on the compute-heavy side.” OpenAI is routing some Dot tasks to Terra, a smaller GPT-6.1 model, instead of GPT-6 Astra, and Huet said Terra can outperform Astra in areas such as computer use and writing at “a tenth of the price.”

Privacy remains the harder problem to route around. Meredith Whittaker warned that personal agents need deep access to user data to act autonomously, raising consent issues for people who email or message an agent user without agreeing to have an AI read their communications.

Huet recommends building trust in stages: connect a calendar, then Slack, then Gmail. His own boundary is blunt: let a Dot categorize email and prepare possible replies, but “don’t send any emails on my behalf ever” and “don’t send any Slack messages on my behalf ever.” As Huet put it, “I want to do that myself.”

OpenAI leadership said users can already see what a Dot accesses and that the company is considering “an easy way to pick and choose” which connections it can use. That control will matter, because a separate Ars OpenForum thread titled “Claude Code settings files expose thousands of secrets on GitHub” described “Do not ask again” prompts leading to passwords and API keys being shared. The thread, started by JournalBot, showed 122 replies from Today at 12:24 PM.

Codex’s next challenge is not adding another feature to the list. It is making local work, cloud work, voice control, persistent agents, and permission boundaries behave like one coherent system. The agents are already busy; now the interfaces need to catch up.

Stay connected via Google News

Follow us for the latest AI News