Karpathy Called Agents Slop. Now He's Running 700 Overnight at Anthropic.
Andrej Karpathy publicly called agentic output 'slop' in October 2025. This week he joined Anthropic to build overnight research loops that run 700 experiments per two-day run — and logged an 11% training speedup. The critique wasn't wrong. The scaffolding was.
NeuroX AI · May 21, 2026
Karpathy called agentic output "slop" on the Dwarkesh Patel podcast in October 2025. This week he joined Anthropic's pre-training team to build exactly that kind of loop — 700 experiments per two-day autonomous run, with an 11% training speedup already on the board.
His open-source autoresearch project — 630 lines, dropped on GitHub in early March — hit 80,000 stars within weeks. The mechanics are straightforward: agents propose changes to training configurations, run the experiments overnight, evaluate the results, and stack the wins. The system found roughly 20 stacking improvements across those early runs. No human in the loop between prompt and outcome.
That's the thing about the "slop" framing. It wasn't wrong about what most agents produce. It was describing what happens when you point an agent at a task without evals, without pre-scoped success criteria, without a feedback loop the system can iterate against. The output reflects the scaffolding, not the model.
The same gap shows up in product engineering. Agents that ship aren't smarter than the ones stuck in staging — they have cleaner inputs, explicit eval gates, and staged environments the agent can actually poke at safely. The model is the easy part. The thirty days around it are the work.