Anthropic adds dreaming

Anthropic’s new “dreaming” feature is not about AI getting weirdly human. It is a memory-maintenance system for agents — the kind of Claude-powered software that runs tasks over time, not just one chat at a time. That matters because long-running agents usually fail in a boring way: they forget, repeat mistakes, and accumulate junk instructions. Anthropic’s pitch is that agents should be able to review their own history between jobs and come back a little sharper next time. (claude.com) ### What did Anthropic actually launch? On May 6, Anthropic added dreaming to Claude Managed Agents as a research preview, alongside broader access to outcomes, multiagent orchestration, and webhooks. Managed Agents is Anthropic’s hosted service for “long-horizon” work — agents that keep going across many steps, tools, and files instead of answering one prompt and stopping. ### What does “dreaming” mean here? Basically, it is a scheduled cleanup-and-learning pass. (claude.com) The system reviews prior agent sessions and memory stores, looks for patterns, extracts useful lessons, removes stale or duplicate entries, and rewrites memory into something more usable. Anthropic says that lets agents carry forward recurring preferences, successful workflows, and lessons from repeated failures. ### Why isn’t normal memory enough? (claude.com) Because raw memory gets messy fast. If an agent stores every useful-seeming note from every session, the memory bank turns into a junk drawer — full of duplicates, outdated assumptions, and half-right rules. Dreaming is Anthropic’s answer to that problem: memory captures what happened during work, then dreaming refines that memory after the fact so it stays high-signal. ### Why does this matter for enterprise agents? (claude.com) Enterprise buyers do not just want a clever demo. They want agents that can handle repetitive workflows with less babysitting. Anthropic is clearly aiming at that market: long-running work, shared preferences across teams, and multiagent setups where several agents need to benefit from the same accumulated lessons. That is the difference between “helpful assistant” and “software worker.” ### What is the control point? The important part is that Anthropic did not make this fully automatic by default. Developers can let dreaming update memory on its own, or they can review proposed changes before anything lands. That fits Anthropic’s broader line on trustworthy agents — keep humans in control, make actions legible, and avoid giving autonomous systems too much unchecked freedom. ### How does this connect to the other launch? (claude.com) Dreaming shipped with “outcomes,” which lets developers define a success rubric and have a separate grader judge the agent’s output. If the work misses the bar, the agent gets another pass. Anthropic says outcomes improved task success by up to 10 points in testing, with gains of 8.4% on docx generation and 10.1% on pptx generation in internal benchmarks. Put together, outcomes handles within-run correction, while dreaming handles between-run learning. ### What is the catch? Self-improving agents sound great, but they also create a new failure mode: an agent can learn the wrong lesson and spread it across future tasks. Anthropic has been pretty explicit that agents already raise risks around misread intent, unintended actions, and prompt-injection attacks. So dreaming is useful only if the memory it produces stays auditable and reversible. ### Bottom line? Anthropic is trying to turn agent memory from a passive notebook into an active feedback loop. (claude.com) That is a real step forward — but the bigger story is not the name “dreaming.” It is that AI companies are now building the plumbing for agents that improve over time, and the whole game will be whether those improvements stay reliable enough for real work. (anthropic.com)