Anthropic Launches Managed Agents: Dreaming and Outcomes
Anthropic shipped three new Claude Managed Agent capabilities at "Code with Claude 2026" in San Francisco. Dreaming (research preview) reviews past agent sessions, extracts recurring patterns, and curates persistent memory so agents self-improve over time. Outcomes (public beta) lets users define a success rubric; a separate grader evaluates output and the agent iterates until the bar is met — reportedly +10 points on hardest task sets. Multiagent orchestration (public beta) enables a lead agent to delegate to parallel specialist sub-agents with webhook notifications on completion.
Why It Matters
These capabilities move Claude from a stateless API to a persistently learning agent platform, closing the gap with open-harness memory frameworks and enabling autonomous overnight enterprise workflows at a first-party platform level.