Anthropic expands Claude Managed Agents platform with Dreaming feature to help AI agents learn from past sessions and improve reliability

Anthropic introduced three new features for its Claude Managed Agents platform (launched in April): Dreaming, Outcomes, and Multiagent Orchestration. Dreaming reviews past agent sessions to spot patterns and share insights across sessions, such as recurring errors or proven workflows.

Dreaming works as an asynchronous job that reads an existing memory store and optionally up to 100 past sessions, removes duplicates and outdated entries, and builds a new organized memory while keeping the original memory intact. Claude Opus 4.7 and Claude Sonnet 4.6 are currently supported, with billing following standard API token pricing.

Outcomes moves from research preview to public beta and lets developers define success criteria (a rubric); a separate evaluator then checks the agent's output against these criteria in its own context window, and the agent can revise its work up to three times by default, with a maximum of 20 attempts if the result does not meet the specs.

Multiagent orchestration (moving to public beta) allows a lead agent (coordinator) to manage several specialized agents running in parallel in separate threads with isolated contexts, up to 20 different agents and a maximum of 25 threads running simultaneously. Dreaming is available as a research preview; developers can request access through a form on the Claude website.