Anthropic announced "dreaming" for Claude Managed Agents on May 6 at its Code with Claude conference. Despite the evocative name, the mechanism is straightforward: a scheduled background process reviews an agent's prior session logs and memory store, merges duplicates, removes stale entries, and surfaces recurring patterns (repeated mistakes, team preferences, recurring task structures) so the agent's persistent context improves over time. Importantly, this is not changes to the model itself — it's curated memory updates to the agent's notes. You can have it run automatically or require human review before changes land. Early adopter Harvey (legal AI) reports task completion rates increased roughly 6× after implementation; Anthropic shipped two companion features the same day — "outcomes" (a self-grading loop where an evaluator scores agent outputs against a rubric) and multiagent orchestration (a lead agent fans work out to specialist sub-agents). Dreaming remains in research preview; access by request.
The implications are worth sitting with. Most agent systems today have terrible long-term memory hygiene — they accumulate context that contradicts itself, stale references that mislead, and patterns the agent should have learned that simply never get encoded. Dreaming addresses that at the agent layer rather than the model layer, which is the practical place to address it. If you are building or operating Claude-based agents at scale, this is the feature to watch. The risk is the obvious one: between-session memory updates without good guardrails are exactly the surface where an agent drifts in directions you cannot see. Anthropic's optional human-review pathway is the right design for production use; treat it as required, not optional, in regulated contexts.