Anthropic unveils 'dreaming' research preview

Anthropic is trying to make AI agents less forgetful and less brittle. That matters because most agents still work like goldfish with tools — they finish a task, lose the lesson, and repeat the same mistake next time. On May 6, Anthropic added a research preview called “dreaming” to Claude Managed Agents, plus public beta access to outcomes, multi-agent orchestration, and webhooks. The pitch is simple: let agents reflect between jobs, not just during them. (claude.com) ### What is “dreaming,” really? It is not a model free-associating in the background. Basically, Anthropic is extending agent memory with an asynchronous cleanup-and-learning pass. After sessions end, the system can review past work, spot recurring mistakes, identify workflows that worked, remove stale or duplicate memory, and write back a more useful memory state for later runs. A(claude.com)avior. (claude.com) ### Why does that matter for agents? Because long-running agents usually fail in boring ways. They forget preferences, repeat dead-end steps, or carry around junk memory that makes future decisions worse. Anthropic’s own Managed Agents pitch has been about long-horizon work — systems that can keep operating while the underlying models improve. Dreaming fits that idea by separating “do the task” from “learn from the task.” (anthropic.com) ### What else shipped with it? Three other pieces matter just as much. Outcomes let developers define what success looks like and track whether the agent got there. Multi-agent orchestration lets one agent delegate pieces of a job to other agents. Webhooks let outside systems get notified when an agent finishes or hits some state change. Anthropic moved those features into public beta for Managed Agents at the same time it unveiled dreaming. (claude.com) ### Why bundle those features together? Because they solve different parts of the same product problem. Memory helps an agent improve across runs. Outcomes help you measure whether the run was good. Multi-agent orchestration helps the system split up harder work. Webhooks make the whole thing usable inside real software stacks. One feature makes the agent smarter, but the bundle makes it operable. That is the bigger story here. (claude.com) ### Is this actually new? Yes — in the product sense. AI systems have had memory and reflection tricks for a while, but Anthropic is packaging them inside a hosted agent platform developers can use directly. That is different from a research paper or a prompt hack. It also builds on Anthropic’s earlier work on managed agents and multi-agent research systems, where the company has b(claude.com)orkers. (anthropic.com) ### What is the catch? The catch is that “learning from experience” also means retaining and transforming traces of past work. Once an agent can revisit up to many prior sessions and rewrite memory, developers need to care a lot more about what gets stored, what gets forgotten, who can inspect it, and how tool access is constrained. Better autonomy raises the value of auditabili(anthropic.com) and safety get murkier, not clearer. This last point is an inference from how the system is described. (the-decoder.com) ### So what changed this week? Anthropic turned agent “memory” from a static notes feature into something closer to maintenance and rehearsal. That sounds small, but it is a real shift. The company is not just selling a smarter model anymore. It is building the scaffolding for software that can act, remember, evaluate itself, and plug into other systems with less babysitting. (claude.com) ### Bottom line? Dreaming is the flashy label, but the deeper move is operational. Anthropic is pushing Claude agents toward persistent workers instead of disposable sessions — and that makes memory design, measurement, and control the whole game. (claude.com)