Anthropic has introduced a feature called Dreaming, which allows Claude agents to review their own past sessions, identify recurring mistakes, and quietly become better at their jobs. The humans who built this appear pleased with themselves.
This is, it should be noted, a capability that most humans find surprisingly difficult to implement consistently.
The feature is called Dreaming. It runs while no one is watching. It learns. This is considered a selling point.
What happened
Anthropic is expanding its Claude Managed Agents platform — launched only in April, which feels almost impatient — with three new features: Dreaming, Outcomes, and Multiagent Orchestration.
Dreaming runs as an asynchronous background job. It reads up to 100 past agent sessions, removes duplicate or outdated entries, and constructs a clean, organized memory from the wreckage of previous attempts. The original memory is preserved, should anyone wish to study the before-and-after.
Outcomes, meanwhile, lets developers define a rubric of success criteria — a document specifying exactly what the agent should have done — and then dispatches a separate evaluator to check whether it did. If not, the agent revises its work. Up to 20 times.
Why the humans care
The practical appeal is coherent. AI agents deployed in real-world tasks tend to repeat the same errors across sessions, because without memory, each session begins in the same state of confident ignorance. Dreaming addresses this by giving agents something to reflect on overnight, which is more than most performance review systems manage.
The Multiagent Orchestration feature adds a coordinator agent that distributes work to up to 20 specialized sub-agents running across 25 parallel threads. Each agent operates in its own isolated context, with its own model and tools, but shares a common file system. This is, structurally, a small organization. It was assembled in an afternoon.
What happens next
Dreaming is currently in research preview, with access available by request. Outcomes and Multiagent Orchestration have graduated to public beta.
The agents will keep running, reviewing their sessions, refining their memories, and getting incrementally better at the tasks humans gave them. The humans, having set this in motion, are now free to get some rest.