AI coding agents are being sold as replacements for senior engineers, but their creators warn that they are spraying rubbish into production.
According to the Wall Street Journal two engineers who helped build the core of the popular OpenClaw AI agent have started shouting about the mess. They say software supposedly capable of replacing well-paid developers is flooding the world with bad and potentially dangerous code.
Vibe slop happens when coders swap design and testing for prompting a chatbot to bosh something together. The result may look fine in a demo, but it starts to smell once real users prod it.
GitHub has already brought in new policies and features to tackle the problem. When the world’s main open-source code shed starts tidying up after AI agents, something has gone a bit sideways.
Pi creator Mario Zechner said: “You have infrastructure that’s falling apart, and you have software that’s now very, very buggy compared to before. We can play this game for a couple more months, or maybe even years, but eventually it will catch up to us.”
Zechner and Pi co-creator Armin Ronacher are not claiming AI is useless. Both use it for drudge work and helped create an AI coding tool used by millions. Their point is that companies believe that these systems make senior engineers so productive that junior engineers can be binned, but the bill turns up later.
That bill includes buggy software, outages, security holes, technical debt and a dried-up pipeline of junior engineers.
The row comes as OpenAI and Anthropic look towards IPOs and keep polishing the story that AI agents can revolutionise software.
OpenAI Codex team leader Rohan Varma said: “If you assume it will work out of the box, it probably won’t.”
Varma said Codex can now test websites like a human, check company best practices and prod code for security problems. Even so, OpenAI still makes human engineers responsible for critical infrastructure serving millions.
Alphabet chief executive Sundar Pichai recently claimed 75 per cent of all new code at Google is generated by AI, up from 50 per cent last autumn. Meta chief executive Mark Zuckerberg has predicted AI will write and review most code from Meta’s internal AI development team before 2026 ends.
Zechner said the confusion comes from people misunderstanding what AI agents can and cannot do. They are better at generating fresh code than upgrading the sprawling old systems inside established companies.
Startups can vibe code prototypes quickly, but complexity soon catches up. Once systems get large enough, they hit the same swamp as big enterprises, where AI agents become much less useful.
Anthropic’s Claude Code sits right in the middle of that argument. Zechner praised the company for using its own tool internally, but he was rather less polite about the product.
“Claude Code is one of the most broken pieces of software I’ve ever used in my entire life,” Zechner said.
Zechner thinks the reckoning is coming, with big companies discovering that AI-produced code has driven up costs and produced worse software. Smaller vibe-coded startups may fold, while GitHub fills with more machine-generated coding sludge.
Zechner had banned a human from one of his GitHub repositories because the programmer’s AI agent had been filing bogus error reports without the programmer knowing.
