George Hotz: AI agents will degrade software quality

George Hotz published a blog post Sunday warning that widespread use of AI coding agents will lower average software quality after six months of testing the tools on real projects.

Hotz wrote the post, titled “The Eternal Sloptember,” after six months using agent tools on parts of Tinygrad, his open-source deep learning framework, and during a full firmware reverse-engineering of a USB-to-PCIe chip.

He wrote that agents “frontload all the progress” but often fail to deliver reliable, correct implementations, leaving engineers to complete manual fixes the models do not finish.

Hotz described the issue as organizational: experienced engineers read generated code, detect mistakes and calibrate when to trust the tool. He wrote the “bottom performers won’t have that self check” and are producing roughly ten times their previous output with agents.

He wrote that at large companies that mix of skills will lead to faster degradation of average code quality masked by greater output velocity.

Hotz pointed to internal deployments of agent tools at major companies and asked whether large codebases such as macOS will improve or worsen if those tools are used broadly without stronger verification processes.

He aligned his view with critics including Yann LeCun and Gary Marcus, arguing that large language models match existing code patterns but do not reason from first principles. He added, “Agents cannot program, and it’s taking longer and longer to realize that they can’t.”

The post appeared five days after Andrej Karpathy posted on May 19 that he had joined Anthropic’s pre-training team and described the coming years as “especially formative.” Anthropic’s chief executive, Dario Amodei, commented that some engineers at the company have stopped writing code themselves and let models handle it while they review outputs.

Agent-based coding expanded rapidly after 2024. In 2025, Microsoft converted GitHub Copilot into an agentic platform, and other vendors repositioned developer tools around agent workflows.

Proponents say agents can increase productivity by handling routine work and generating drafts. Critics warn the tools can produce subtle errors that are costly if not detected.

Hotz’s post adds a prominent skeptical voice to an industry debate among engineers and researchers about whether agent workflows produce production-quality software.