“In another two to three months, Codex will be a primitive tool.”
The line did not come from a rival. It was attributed to Thibault Sottiaux, OpenAI’s general manager for product and platform, described in the article as the executive responsible for ChatGPT and Codex.
The article says harness frameworks have been breaking out beyond a narrow developer audience. Developers and office workers alike are using agent-style tools and getting productivity gains from AI automation. Within that group, OpenAI’s Codex is presented as one of the standouts.
Even so, Sottiaux’s view is that the next step for AI agents may not be a simple tool upgrade. The change could be more structural, moving from small single-machine utilities to cloud-native infrastructure.
Why local harness setups are hitting limits
The article defines a harness as the layer built around a large model: context management, tool invocation, state persistence, environment isolation, and fault recovery. A large share of developers still run agents on laptops, using setups built around a CLI, Cursor, Claude Code, or lightweight agent logic.
At the same time, multi-tool and multi-model workflows are becoming more common. One example in the piece describes a developer using Fable 5 inside Claude Code to draft project documentation, then switching to Codex to execute work with other models.
That pattern, the article argues, runs into three physical bottlenecks when it meets the next generation of frontier models.
Compute and memory ceilings
If an agent needs to run 20 subtasks in parallel — such as testing, scraping data, and compiling a large project at the same time — a local laptop can quickly exhaust memory, CPU, and concurrent sandbox resources. The piece specifically mentions Docker and virtual machines.
Long-running jobs make local execution awkward
Complex agent tasks may run for hours or even days. That leaves the user’s laptop unable to shut down, disconnect from the network, or even close its lid. The article describes that engineering setup as highly impractical.
Parallel context and tool-chain management becomes harder
The article says next-generation models are likely to support much longer context windows and highly parallel reasoning. Lightweight local harness systems may struggle to manage context compression, state synchronization, and centralized log tracing across large numbers of concurrent agents.
Cloud execution is already entering the stack
According to the article, signs of a shift toward heavier infrastructure and more system-oriented agent design have been visible over the past few months.
Codex is one example. The piece says OpenAI’s product already supports asynchronous cloud execution. A user issues a command from the terminal, and the job is sent to an isolated cloud container — described in the article as a Sandbox Container — where it runs autonomously. The laptop or phone mainly acts as a control surface.
Another trend is the rise of cloud-native micro-sandbox infrastructure. The article names E2B, Daytona, Fly.io, and Modal as platforms built for AI agents, saying they can spin up hundreds of isolated container environments on demand so agents can run code in parallel and verify results.
Teams are shifting from prompt work to system-level harness design
The article says engineers at OpenAI, Anthropic, and Cognition are beginning to move their effort away from prompt optimization and toward writing system-level harnesses.
It points to a February example from Anthropic. In that demonstration, 16 Claude instances ran in parallel across 2,000 cloud sessions and produced a C compiler. The article describes the project as a milestone for multi-agent collaboration in software engineering.
Different Claude instances handled different roles. One architect agent designed the abstract syntax tree, or AST. Four coding agents worked on separate modules. Two testing agents wrote unit tests. One audit agent reviewed the code flow and security.
Nicholas Carlini, identified in the article as the lead researcher, said: “Most of the effort went into building the environment, the testing loop, and the feedback infrastructure for the models.”
“Light local control, heavy cloud execution”
The article says workflows of this kind could become routine in the next two to three months. In line with Sottiaux’s view, it argues that “light local control, heavy cloud execution” is set to become the default setup, with developer interfaces such as IDEs, terminals, and web UIs turning into control panels.
Under that model, compute-heavy work — code refactoring, test execution, and browser automation among it — would be handled in parallel inside dynamic cloud clusters. Diffs and log streams would then be pushed back to the local machine.
The article closes by arguing that as models and harnesses become more deeply integrated, the competitive focus may shift from raw model reasoning power to how fully that power can be unlocked. Agent frameworks that still rely mainly on local Python scripts and simple API calls, it says, are nearing their ceiling.
The piece was originally published by the WeChat account Jiqizhixin (ID: almosthuman2014) and credited to the author Zenan.