Thibault Sottiaux, head of the OpenAI platform, said that AI agents are evolving from single-machine tools to cloud-native infrastructure.Article author, source: MachineHeart
In another 2-3 months, Codex will be an outdated tool.
This statement was not made by a competitor—it was made by Thibault Sottiaux, the current General Manager of Products and Platform at OpenAI, who oversees ChatGPT and Codex.
Recently, the Harness framework has gained widespread adoption, with users ranging from developers to office workers embracing this AI agent tool and reaping the benefits of automation. Among these, OpenAI’s Codex stands out as a leader.
However, at this juncture, OpenAI’s executives have already looked further ahead. The next step for AI agents may involve a paradigm shift from standalone tools to a “cloud-native infrastructure.”
What Thibault Sottiaux meant by “the laptop is no longer enough” is clear: we are trying to run next-generation models with long-context reasoning and high autonomy on single-machine Harness setups—a mismatch akin to putting big feet into small shoes—that can no longer be sustained.
We understand that Harness refers to a suite of features built around large models, including context management, tool invocation, state persistence, environment isolation, and error recovery. Currently, many developers accomplish tasks by running agents on their laptops—such as through CLI, Cursor, Claude Code, or lightweight agent logic.
An increasing number of developers also prefer using multiple tools and models together—for example, using Fable 5 on Claude Code to write project documentation, then switching to Codex to execute tasks with another model.
But this encounters three major physical bottlenecks when facing next-generation frontier models:
Computing power and memory limitations: When an agent needs to concurrently execute 20 subtasks—such as running tests, scraping data, and compiling a large project—the local laptop’s memory, CPU, and concurrent sandbox resources (Docker/VM) will be instantly overwhelmed.
Long-running tasks cannot be shut down: Complex Agent tasks may require hours or even days to complete, forcing users to keep their laptops powered on, connected to the internet, and open—this is highly impractical from an engineering perspective.
Parallel explosion of context and toolchains: Next-generation models will certainly support extremely long contexts and highly concurrent inference. Local lightweight harnesses will struggle to handle context compression, state synchronization, and centralized logging across large-scale concurrent agents.
In response to these challenges, signs of Agent tools transitioning toward heavy infrastructure and system design have gradually emerged over the past few months.
For example, OpenAI’s Codex currently supports cloud-based asynchronous execution. After users issue commands in the terminal, tasks are sent to isolated cloud containers (sandbox containers) to run autonomously, while mobile devices or laptops serve only as control interfaces.
Meanwhile, cloud-native micro-sandboxes are emerging: cloud-based, ultra-fast sandbox infrastructures like E2B, Daytona, Fly.io, and Modal—designed specifically for AI agents—enable agents to dynamically spin up hundreds of isolated container environments to run code and verify results in parallel.
Among top AI teams including OpenAI, Anthropic, and Cognition, engineers are beginning to shift their R&D focus from optimizing prompts to “building system-level harnesses.”
In February this year, the Anthropic team demonstrated a case in which 16 Claude instances ran in parallel across 2,000 cloud sessions to write a C language compiler. This marked a milestone in generative AI’s advancement toward multi-agent collaboration in software engineering. Different Claude instances played distinct roles: one architect agent designed the abstract syntax tree (AST), four coding agents handled different modules, two testing agents wrote unit tests, and one audit agent reviewed the code flow and security.
Lead researcher Nicholas Carlini noted: "Most of the effort has been focused on building the environment, testing the closed-loop, and feedback infrastructure for the model."
In the next two to three months, such agent workflows may become part of our daily routine.
Combined with Thibault Sottiaux’s assessment, “local light control, cloud-heavy execution” will become standard, and developer interfaces (such as IDEs, terminals, and web UIs) will fully transform into control panels. Computationally intensive tasks—such as actual code refactoring, running tests, and browser automation simulations—will all be processed in parallel on dynamic cloud clusters, with final diffs and logs streamed back to the local environment.
The deep integration of models with Harness may shift the competition from “whose model has stronger reasoning capabilities” to “whose model’s potential is fully unleashed.” Agent frameworks that rely solely on running Python scripts and simple API calls on local computers are about to hit their limits.