Anthropic launched a new set of Claude agent features Wednesday, including a dreaming tool designed to help its AI agents remember past work, spot recurring mistakes, and improve over time.
Dreaming is designed to extract patterns, curate memories, and surface recurring mistakes or workflows that individual agents may not identify on their own. Developers can let dreaming update memory automatically or review changes before they are added.
The company said the feature is especially useful for long-running work and multiagent orchestration, where agents need to retain useful context across sessions. Memory allows each agent to capture what it learns while working, while dreaming refines that memory between sessions and pulls shared learnings across agents.
Anthropic also made outcomes, multiagent orchestration, and webhooks available to developers building with Managed Agents. Outcomes allow developers to write rubrics that define success, with a separate grader evaluating an agent’s output and prompting revisions when the result falls short.
In internal testing, Anthropic said outcomes improved task success by up to 10 percentage points over a standard prompting loop, with the largest gains on harder problems. The company also said outcomes improved file generation quality, increasing task success by 8.4% for docx files and 10.1% for pptx files in internal benchmarks.
Multiagent orchestration allows a lead agent to divide complex tasks among specialist agents, each with its own model, prompt, and tools. The agents can work in parallel on a shared filesystem, contribute to the lead agent’s context, and remain traceable through the Claude Console.
Anthropic said teams are already using the tools to build agents that verify their own work, learn across sessions, and handle complex workflows in parallel. Harvey used dreaming to help legal agents remember filetype workarounds and tool specific patterns, with completion rates rising about six times in tests.
Netflix’s platform team used multiagent orchestration to analyze logs from hundreds of builds and identify recurring issues across applications.
The update expands Anthropic’s push to make Claude agents more useful for enterprise work where accuracy, memory, and task coordination matter. Dreaming remains in research preview, while outcomes, multiagent orchestration, and memory are available in public beta as part of Managed Agents.
