Anthropic Launches 'Dreaming' AI Agents That Self-Improve Over Time With New Multiagent Orchestration
Summary
Anthropic launches 'dreaming' AI agents that self-improve by reviewing past sessions, introduces outcome-based self-correction showing up to 10-point task success gains, and rolls out multiagent orchestration allowing lead agents to delegate complex tasks to specialized subagents — with early adopters like Netflix and Harvey already reporting significant performance improvements.
Key Points
- Anthropic is launching 'dreaming' in Claude Managed Agents as a research preview, a scheduled process that reviews past agent sessions to extract patterns, refine memory, and enable agents to self-improve over time without constant human oversight.
- New 'outcomes' functionality allows developers to define success rubrics so agents can evaluate and self-correct their own outputs using a separate grader, with internal benchmarks showing up to 10-point improvements in task success and notable gains in file generation quality.
- Multiagent orchestration is now available, enabling a lead agent to break complex tasks into parallel workstreams handled by specialized subagents, with full traceability in the Claude Console — early adopters like Harvey, Netflix, and Spiral by Every are already reporting significant performance gains.