Agentic AI hits a wall of real-world friction
Zuckerberg publicly admitted that Meta's AI agent development is going slower than expected, prompting a thread full of skepticism and some pointed jokes about who told him it would be fast. Separately, a controlled study on Claude Code across 660 trials found that messy codebases don't hurt agent pass rates but do substantially inflate the operational footprint, meaning agents burn more tokens and take more steps in dirty repos.
The pattern here: two completely separate threads are circling the same thing. Agents work in demos and in clean conditions, but real deployment is messier, slower, and more expensive than advertised. OpenAI simultaneously announced GPT-5.6 Sol Ultra coming to Codex with a 'subagents' mode for complex work, which generated questions about whether individual subscribers would even get access.
Commenters on the code cleanliness study pushed back on one of the study's methods, arguing that auto-cleaning messy repos before running agents is a bad approach that papers over real architectural problems. The nuance matters: throwing agents at bad code might 'work' on paper but creates a different category of technical debt.
So what?
If you're building on top of agentic AI, your infrastructure costs are directly tied to the state of your codebase, not just model pricing. Clean code is now an operational budget concern, not just an aesthetic one. And if Meta with unlimited resources is struggling with agent reliability, be very careful about what you promise customers on timelines.