Anthropic gives Claude agents dreaming to learn between runs

Agents usually reset every run, so any improvement loop had to be built manually. Anthropic is baking that in with dreaming, a scheduled pass where Claude agents go back over past sessions, pull out patterns, and update what they remember about a user’s workflow.

The rollout pairs that with outcomes, multiagent orchestration, and webhooks in public beta, while dreaming sits behind a waitlist as a research preview. The mechanism is memory consolidation: spotting recurring mistakes and preferences between runs, then curating or updating memory so the next session starts smarter instead of fresh.