Anthropic introduced "dreaming," a new capability that lets Claude-powered agents learn from their own past sessions and self-improve between tasks. The feature was unveiled at the company's Code with Claude developer conference on May 6, 2026, and is one of the most concrete attempts yet to ship a continual-learning loop inside a commercial agent platform.
The move arrives as enterprises increasingly demand AI agents they can trust with production workloads. Anthropic's pitch: rather than treating each agent run as a stateless task, give the agent the equivalent of sleep — time to reflect, consolidate memory, and improve.
How dreaming works
According to Anthropic, dreaming is a scheduled process that allows agents to "review earlier sessions and their memory stores, extract patterns from them, and then curate memories that could be useful in future." The system surfaces patterns no single agent run can see on its own, including recurring mistakes, workflows that multiple agents independently converge on, and preferences shared across a team of agents.
The feature is currently in research preview, meaning developers must request access before integrating it into their deployments.
Outcomes and multi-agent orchestration
Dreaming was announced alongside two other capabilities for Claude Managed Agents, both now in public beta:
- Outcomes uses a separate "grader agent" to evaluate outputs against ideal examples. Anthropic says internal tests show task success improves by up to 10 points compared with standard prompting.
- Multi-agent orchestration lets a managed agent break down a complex task and assign pieces to sub-agents, with visibility into each sub-agent's work through the Claude Console.
Anthropic also doubled Claude Code's five-hour rate limits for Pro and Max users — and removed peak-hour throttling — addressing complaints from heavy Claude Code users. The capacity bump is enabled by a new compute deal with SpaceX for the Colossus 1 data center.
Early customer signals
Coverage of the launch by VentureBeat cited several early production results: legal AI company Harvey reported task completion rates rising roughly 6x after deploying dreaming, while medical document review firm Wisedocs cut review time by 50% using outcomes. Netflix is reportedly using multi-agent orchestration to process logs from hundreds of builds in parallel.
Those figures are vendor-supplied and have not been independently verified, but they hint at why enterprise buyers are treating self-improving agents as the next meaningful step beyond static chatbots.
Why it matters
Dreaming is small in scope today — a research preview gated behind access requests — but architecturally it points at where the agent stack is heading. Most production agents are forgetful: they finish a task, drop their context, and start the next session naive to what just happened. Persistent, curated memory turns an agent from a tool that runs prompts into a worker that accumulates institutional knowledge.
For enterprises, that changes the buying conversation. Procurement teams that were comparing vendors on benchmark scores will increasingly ask about memory hygiene, drift control, and how an agent's behavior is supposed to evolve over weeks of deployment. Anthropic's bet is that owning the memory layer — not just the model — is what locks in long-term agent customers.
