Anthropic Unveils 'Dreaming': Agents That Review Past Sessions and Self-Improve Overnight

simonwillison.net·high signal

At Code with Claude, Anthropic announced a research preview of 'Dreaming' — a capability that allows agents to review previous sessions, identify gaps in their approach, and self-improve through overnight processing. This represents a shift toward persistent, self-correcting AI agents that learn across sessions rather than starting fresh each time. Combined with the new 'Outcomes' feature (agents iterating toward defined success criteria), Anthropic is building toward agents that compound their effectiveness over time.