Anthropic enriches Claude Managed Agents with Dreaming, Outcomes, and multi-agent orchestration.

Anthropic upgrades Claude Managed Agents with Dreaming and Outcomes to boost autonomous agent stability and reduce human oversight on complex tasks.

At Code with Claude, Anthropic unveiled several additions to Claude Managed Agents, its hosted autonomous agent platform. Four features are being introduced to make agents more stable over time and limit human oversight on long or complex tasks.

The most novel feature, dubbed Dreaming, is available as a research preview by request. The mechanism involves reviewing the agent's past sessions and its memory store, extracting recurring patterns (repeated errors, efficient workflows, team preferences), eliminating duplicates and contradictions, then restructuring everything. Memory can be updated automatically or submitted for manual validation. Harvey, a developer of a legal assistant, reports a sixfold increase in its completion rate after activation.

Outcomes, in public beta, enables the definition