Anthropic launches 'dreaming' so Claude agents learn from their own mistakes

At its second Code with Claude conference, Anthropic introduced 'dreaming' — a system where Claude Managed Agents replay past sessions, identify patterns in their own failures, and self-improve between runs. Anthropic also moved 'outcomes' (developer-defined success metrics) and multi-agent orchestration from research preview to public beta. Mercado Libre showcased Claude Code running across 23,000 engineers, targeting 90% autonomous coding by Q3.