"In another 2-3 months, Codex will be a primitive tool."
This wasn't said by a competitor, but by Thibault Sottiaux, the current OpenAI General Manager of Product and Platform responsible for ChatGPT and Codex, who came forward and said it himself.
Recently, the Harness framework has gradually broken out of its niche. From developers to white-collar workers, people are widely adopting these Agent tools, enjoying the automation benefits brought by AI. Among these, OpenAI's Codex stands out as a top performer.
Yet at this juncture, OpenAI executives are already looking further ahead. The next step for AI Agents might involve another paradigm shift evolving from single-machine tools towards "cloud-native infrastructure."
The meaning behind Thibault Sottiaux's statement that "notebooks aren't enough anymore" is clear: we are using single-machine harnesses to drive the next generation of models that possess long-horizon reasoning and high autonomy. This "big shoes on small feet" model is nearing its limits.
As we know, so-called Harness refers to the suite of capabilities built around large models for context management, tool calling, state persistence, environment isolation, and error recovery. Currently, many developers are running Agents (e.g., via CLI, Cursor, Claude Code, or lightweight Agent logic) on their laptops to complete tasks.
Of course, more and more developers are also inclined to use multiple tools and models together: for example, using Fable 5 on Claude Code to write project documentation, then switching to Codex to execute with other models.
However, this approach will encounter three major physical bottlenecks when facing next-generation frontier models:
Compute and Memory Limitations: When an agent needs to execute 20 subtasks concurrently (like running tests, scraping data, and compiling a large project simultaneously), the local notebook's memory, CPU, and concurrent sandbox (Docker/VM) resources will be instantly overwhelmed.
Long-running Tasks Cannot Shut Down: Complex agent tasks may need to run for several hours or even days, requiring the user's notebook not to be shut down, disconnected from the internet, or closed. This is highly impractical from an engineering perspective.
Parallel Explosion of Context and Toolchains: Next-generation models will undoubtedly support extremely long contexts and highly concurrent reasoning. Local lightweight harnesses will struggle to handle context compression, state synchronization, and centralized log tracing among large-scale concurrent agents.
Faced with these challenges, signs of Agent tools transitioning towards heavyweight infrastructure and system design have actually been gradually emerging over the past few months.
For example, OpenAI's Codex currently offers cloud-based asynchronous execution support. After a user issues a command in the terminal, the task is sent to an isolated container (Sandbox Container) in the cloud to run autonomously, with the phone or laptop merely serving a command-and-control function.
Simultaneously, cloud-native micro-sandboxes are emerging: Cloud-based, high-speed sandbox infrastructure specifically designed for AI Agents, like E2B, Daytona, Fly.io, and Modal, can dynamically spin up hundreds of independent container environments for agents 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 "writing system-level harnesses."
In February of this year, the Anthropic team showcased a case where 16 Claude instances running in parallel across 2000 cloud sessions wrote a C language compiler. This is a milestone for generative AI moving towards multi-agent collaboration in software engineering. Different Claude instances played different roles: 1 Architect Agent responsible for designing the Abstract Syntax Tree (AST), 4 Coding Agents responsible for different modules, 2 Testing Agents dedicated to writing unit tests, and 1 Audit Agent responsible for reviewing the code flow and security.
Lead researcher Nicholas Carlini pointed out: "Most of the effort was spent building environments for the model, test loops, and feedback infrastructure."
In the next two to three months, such agent workflows might become our daily routine.
Combined with Thibault Sottiaux's prediction, "lightweight local command, heavyweight cloud execution" will become the standard. Developer interfaces (like IDEs, Terminals, Web UIs) will completely transform into control panels. Truly computation-intensive tasks like code refactoring, test execution, and browser automation simulations will all be processed in parallel within dynamic cloud clusters, with only the final differences (Diffs) and log streams pushed back to the local machine.
The deep integration of models and harnesses might shift the competition from "whose model has stronger reasoning capabilities" to "whose model's potential is unleashed more thoroughly." Those agent frameworks that rely solely on running Python scripts and simple API calls on local computers are about to hit a hard ceiling.
This article is from the WeChat public account "Machine Heart" (ID: almosthuman2014), author: Zenan
