Researcher says literature already contradicted Anthropic's shifting subagent guidance months ago

gerardsans · x · 2026-11-17

Researcher Gerard Sans cites his synthesis of agentic AI evidence (grounding/belief-revision limits, multi-agent error amplification, a persistent 30% misbehaviour floor, benchmark-vs-deployment gap) to argue Anthropic's guidance is self-contradictory: first recommending subagents, then advising against them. He claims the literature established these conclusions months ago and that AI labs' statements shouldn't be taken at face value.

Related event: Anthropic Advises Developers to Fold Subagents into Single-Agent Loops(4 posts)→

Original post →

More from Models

Models channel →