Anthropic unveils dreaming feature

Anthropic is trying to turn AI agents from clever interns into something more like long-running coworkers. The missing piece has been memory — not just remembering facts, but cleaning up bad habits, spotting repeated mistakes, and carrying useful patterns forward. On May 6, Anthropic said Claude Managed Agents can now “dream,” which is its name for a scheduled review process that revisits past sessions and updates memory between tasks. The point is simple: let agents improve inside the product, instead of waiting for humans to rewrite prompts and workflows every time something goes wrong. (claude.com) ### What is “dreaming,” actually? It is not the sci-fi version. Anthropic describes it as a background process that reviews an agent’s session history and memory store, extracts patterns, merges or curates memories, and updates what the agent should carry into future work. Developers can let those memory changes land automatically or require review first. Basically, the agent gets a maintenance cycle between assignments. (claude.com) ### Why wasn’t normal memory enough? Because raw memory gets messy fast. If an agent keeps appending notes from every task, the store fills up with duplicates, stale preferences, contradictions, and one-off noise. Anthropic’s pitch is that dreaming keeps memory “high-signal” by reorganizing it between sessions. That matters more for long-running work, where the failure mode is not one bad answer but a slow buildup of junk context. (claude.com) ### What else launched with it? Anthropic paired dreaming with “outcomes” and multiagent orchestration. Outcomes lets a developer define a rubric for success, then uses a separate grader to judge whether the agent met that bar and send it back for another pass if needed. Anthropic says this improved task success by up to 10 points in internal testing, with gains of 8.4% on docx generation and 10.1%(claude.com) helps agents learn over time, while grading helps them correct themselves right now. (claude.com) ### Why is Anthropic pushing this now? Because the company is making a bigger bet on managed agents as a product category. It launched Managed Agents in public beta in April and framed the service as hosted infrastructure for long-horizon work — with stable interfaces for sessions, harnesses, and sandboxes even as the machinery underneath changes. In plain English, Anthropic wants developers to bui(claude.com)hat hosted layer more valuable because the improvement loop lives inside Anthropic’s system. (anthropic.com) ### Where does the SpaceX deal fit? More capable agents need more compute, and Anthropic used the same event to announce a big new supply of it. The company said it will use the full computing power of SpaceX’s Colossus 1 facility in Memphis, getting 300 megawatts of new capacity within a month. Anthropic also said that extra capacity would support higher Claude Code usage limits, including remo(anthropic.com) infrastructure story are really one story — better agents are only useful if enough customers can run them. (money.usnews.com) ### Why is this a bigger deal than a feature tweak? Because it changes what “agent reliability” means. A normal chatbot can be judged turn by turn. A self-improving agent has to be judged across weeks, with memory changes accumulating in the background. That creates a new risk surface — not just obvious policy violat(money.usnews.com) The catch is that self-improvement is useful only if someone is watching the trajectory, not just the latest output. That is why Anthropic kept dreaming in research preview and gave developers a review option for memory updates. (claude.com) ### So what changed today? Anthropic moved one step closer to agents that do their own upkeep. That sounds small, but it is a real shift. The industry has spent two years teaching models to act. Now it is starting to teach them to revise themselves — inside guardrails, with memory as the battleground. (claude.com)