Anthropic introduces “dreaming,” a way for AI agents to review past behavior between sessions, improving performance over time.
Anthropic’s push for more autonomous AI agents just gained a new process called “dreaming,” aiming to help systems learn between sessions rather than start fresh every time.
Misryoum reports that the company unveiled “dreaming” at its developer conference as part of a broader effort to reshape how software engineering and other knowledge work get handled by AI tools.. Built to feed into Anthropic’s Claude Managed Agents offering. the technique is designed to refine an agent’s memory through structured evaluations that revisit earlier behavior and look for patterns.
In practical terms, the goal is to convert past mistakes and lessons into improved working methods, reducing errors as agents iterate over time.
For now. “dreaming” is being offered as a research preview. meaning access is not open to all developers and participation requires an application.. Anthropic positioned the update as a step toward agents that can keep getting better at their tasks. which could raise their usefulness for businesses paying for tools that need consistent. long-running performance.
That matters for the commercial trajectory of agentic AI: if systems can reliably improve their approach over repeated cycles, the value proposition shifts from “try an assistant” to “deploy an operator.”
The release also lands amid momentum for Anthropic’s products aimed at developer workflows.. Misryoum notes that revenue has picked up in recent months as engineers adopt Claude Code and related offerings that support longer-duration coding projects.. Anthropic is simultaneously working to extend these agent capabilities beyond software work, including areas such as finance and law.
Meanwhile. Anthropic also moved two Managed Agents tools from research preview to public beta. including one that uses rubric-style outcomes to guide agent performance and another that can delegate work across multiple sub-agents.. Together. these changes suggest an emphasis on both steering (so outputs follow expectations) and scaling (so tasks can be divided and handled in parallel).
In the background. Misryoum says the company is exploring the broader direction of agentic systems that manage more of their own lifecycle.. As techniques like “dreaming” mature. businesses will likely focus not only on raw capability. but on whether agents can improve safely and predictably as they accumulate experience.