Anthropic is betting that the next leap in enterprise AI won’t come from bigger models—it will come from agents that learn while they work. At its Code with Claude developer conference in San Francisco, the company unveiled “dreaming,” a new capability that lets Claude agents review their own past sessions, identify patterns, and write playbooks to improve future performance.
Unlike simple memory features that recall user preferences within a chat, dreaming operates at a higher level. It is a scheduled process that sifts through an agent’s history, extracts recurring mistakes, notes workflows that multiple agents independently converged on, and surfaces team-wide preferences. The result is a self-correcting system that gets sharper over time without human hand-holding.
Crucially, dreaming does not alter the underlying model weights. Instead, it writes plain-text notes and structured “playbooks” that future sessions can reference. That makes the process fully inspectable and auditable—a key concern for enterprises wary of black-box automation. “They’re learning to write better notes for their future self,” said Alex Albert, Anthropic’s research product lead.
The company also moved two experimental features—outcomes and multi-agent orchestration—into public beta. Outcomes let developers define success criteria for complex tasks, while multi-agent orchestration allows fleets of Claude agents to divide and conquer large projects. Early adopters are already seeing dramatic returns. Legal-tech firm Harvey reported roughly 6× task-completion gains after implementing dreaming. Medical-document startup Wisedocs cut review times by 50 percent using outcomes. Netflix, meanwhile, is now processing logs from hundreds of builds simultaneously with multi-agent orchestration.
The announcements come amid explosive growth for Anthropic. CEO Dario Amodei revealed that first-quarter revenue and usage grew at an 80× annualized clip—far above the 10× the company had planned for. Average developers using Claude Code now spend 20 hours per week with the tool. With dreaming, Anthropic hopes to convert that momentum into autonomous systems that enterprises can trust at scale.
Why It Matters
Self-improving agents could dramatically reduce the human oversight needed for complex enterprise automation, accelerating trustworthy AI adoption across regulated industries.