Claude’s latest AI trick sounds weirdly human

AI companies keep finding new ways to make their chatbots sound unnervingly human, and Anthropic has just added a particularly strange one to the list. Claude agents can now “dream” about their past work, although this is less sci-fi than it sounds.

Anthropic just announced the new dreaming feature for Claude Managed Agents on its website. The idea is that, between active work sessions, a scheduled review process can potentially help agents improve before the next session. It can spot recurring mistakes, identify workflows that agents tend to settle on, and tidy up memory so it doesn’t become a messy pile of notes. Developers can let these memory updates happen automatically or review them before they take effect.

This is aimed at developers building with Claude Managed Agents rather than regular Claude users chatting in the app. It could matter most for long-running work, where an AI agent needs to keep improving across multiple sessions instead of starting from scratch each time.

Anthropic also announced a few related upgrades. Outcomes lets developers define what success looks like for a task, then has a separate grader check whether the agent’s work meets that bar. If it doesn’t, the agent can try again. Multiagent orchestration lets one lead agent split a bigger job between specialist agents, such as having different agents check logs, support tickets, metrics, and so on.

Anthropic says high-profile teams, including Netflix and Wisedocs, are already putting these tools to work. Netflix’s platform team is using them to analyze build logs at scale, while Wisedocs says its document reviews now run 50% faster.

Dreaming is available now as a research preview in Claude Managed Agents, while outcomes and multiagent orchestration are in public beta. Whether it’s disappointing or a relief to find that Claude isn’t drifting off into some digital nap and dreaming about electric sheep will depend on your own view of AI.