Claude Adds Self-Improving Agent Features

dreaming periodically reviews agent sessions and the memory store to find repeated patterns and refine memory, so agents get better over time. It can update memory automatically, or you can choose to review changes before they're applied.

This feature identifies recurring mistakes, workflows that multiple agents have converged on, and preferences shared across a team. It restructures memory so that only useful signal remains, tailoring it for long-running tasks and multiagent orchestration.

Also released alongside it, outcomes defines success criteria as grading criteria, and a separate grader evaluates results in an independent context window. If the criteria aren't met, the grader points out what needs fixing and the agent tries again.

Also added is a flow where you define an outcome and receive a webhook notification when the agent finishes. Internal testing reported up to a 10-point performance improvement over a standard prompting loop, along with task success improvements of +8.4% for docx and +10.1% for pptx.

In multiagent orchestration, a lead agent breaks down tasks and distributes them to specialized subagents. Each subagent uses its own model, prompt, and tools, works in parallel on a shared filesystem, and maintains a history of work through persistent events. In the Claude Console, you can track which agent did what, when, and why.

As examples, Harvey applies dreaming to legal document work, Netflix uses multiagent orchestration for large-scale build log analysis, Spiral applies outcomes to a combination of a Haiku lead agent and Opus subagents, and Wisedocs applies outcomes to quality control for document verification.

dreaming is available as a research preview under Managed Agents, while outcomes, multiagent orchestration, and memory are released in public beta. Developers can request access.