Why Codex-Style Harnesses Will Peak in Just Two Months—and Laptops Won’t Keep Up

OpenAI product and platform GM Thibault Sottiaux, who oversees ChatGPT and Codex, predicts that in the next 2‑3 months Codex‑style Harness agents will become a “primitive tool.” He emphasizes that running agents on a laptop is no longer sufficient for the next generation of models.

The article explains that Harness refers to a suite of components around large models—context management, tool invocation, state persistence, environment isolation, and fault recovery. Most developers currently execute agents locally via CLI, Cursor, Claude Code, or lightweight logic on their laptops.

Three major physical bottlenecks arise when scaling agents:

Compute and memory limits: concurrent execution of 20 sub‑tasks (e.g., testing, data crawling, large‑project compilation) quickly exhausts a laptop’s CPU, RAM, and sandbox resources.

Long‑running tasks cannot be shut down: complex agents may need to run for hours or days, requiring the laptop to stay on, connected, and open.

Context and toolchain parallel explosion: future models will support very long contexts and high‑throughput inference, making it hard for lightweight local Harnesses to handle massive context compression, state synchronization, and centralized logging.

These challenges have already spurred a transition toward cloud‑native infrastructure. OpenAI’s Codex now offers asynchronous cloud execution, where commands are dispatched to isolated sandbox containers and the local device acts only as a control panel. Emerging micro‑sandbox platforms such as E2B, Daytona, Fly.io, and Modal provide rapid, isolated container environments that can spin up hundreds of parallel instances for agents.

Among leading AI teams, engineers are shifting focus from prompt optimization to building system‑level Harnesses. In February, Anthropic demonstrated a multi‑Claude workflow that used 16 Claude instances across 2,000 cloud sessions to compile a C language compiler, assigning distinct roles (architect, code generators, test writers, audit agent) to each instance.

Researcher Nicholas Carlini notes that “most effort is spent building the environment, testing loops, and feedback infrastructure for models.” Looking ahead, Sottiaux envisions a standard pattern of “local lightweight control, cloud‑heavy execution,” where IDEs, terminals, and web UIs become pure control panels while compute‑intensive tasks—code refactoring, test execution, browser automation—run in parallel cloud clusters, streaming diffs and logs back to the user.

The deep integration of models and Harnesses may shift competition from raw model inference power to how thoroughly a model’s potential is unlocked, rendering purely local Python‑script agents obsolete.