Anthropic brings dreaming to Managed Agents

Anthropic has added dreaming to Managed Agents as a research preview.

This feature is a scheduled memory process that reviews recent sessions and work patterns to refresh memory. Execution can be set to fully automated, or configured so that changes are applied only after user review.

Outcomes, which also entered public beta together, works by first defining task criteria, then having a separate grader agent evaluate the results. Anthropic explained that this approach is especially useful for tasks where detailed accuracy and exhaustive coverage matter, as well as tasks requiring subjective quality such as brand tone.

In internal testing, applying outcomes improved task success rates by up to 10 points compared to a standard prompting loop.

Another update is multi-agent orchestration. A lead agent splits up tasks and distributes them to subagents, and in the Claude Console you can check each agent's execution process step by step.

Currently, outcomes and multi-agent orchestration are in public beta, while dreaming is accessible by request.