Agents as Webs of Beliefs

Researchers are proposing a new way to understand intelligent agents: not as unified rational actors, but as webs of beliefs. According to Richard Ngo on LessWrong, an agent's beliefs don't need perfect internal consistency—they just need to cohere locally, like a spider web under tension. The framework pulls together ideas from active inference and agent foundations, unifying beliefs, goals, and actions into a single model. Here's the key insight: actions are actually a special kind of belief—specifically, beliefs about what you'll do that make themselves true. When you believe you'll raise your arm, your nervous system reads that belief and acts on it. This extends to social interactions: if others believe you'll win and help you because of it, then holding that belief is itself an action. The open challenge is formalizing this with the hierarchical, iterative learning we see in deep networks—how belief systems can build concepts from simpler concepts without collapsing into paradox.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton