Claude Managed Agents Gain Dreaming for Self-Improvement & Better Outcomes

TL;DR

• Dreaming enables agents to review past sessions and refine memory, allowing for long-term self-improvement.
• Outcomes introduce a grading system that improved task success by up to 10 points and increased file generation quality in benchmarks.
• Multiagent orchestration allows developers to delegate complex tasks to specialist agents that operate in parallel.

Anthropic has introduced a suite of powerful features for its Managed Agents Overview, fundamentally changing how developers build autonomous workflows. The highlight is the new dreaming capability, currently in research preview. This feature acts as a maintenance layer that reviews agent memory between sessions to identify recurring patterns, team preferences, and workflows. By synthesizing these insights, dreaming ensures that agents become more effective over time without requiring constant manual adjustment.

To address the challenge of quality control, Anthropic has implemented **Defining Outcomes for Agents](https://platform.claude.com/docs/en/managed-agents/define-outcomes). Instead of relying solely on prompt instructions, developers can now provide a clear success rubric. A dedicated grader then evaluates the agent's work against these criteria, enabling the agent to self-correct automatically if the output falls short. This approach has proven highly effective, yielding a +10.1% improvement in PPTX generation quality and significant gains on complex, multi-step problems during internal testing.

When individual agents are insufficient for a task, developers can now utilize **Multiagent Orchestration](https://platform.claude.com/docs/en/managed-agents/multi-agent). This system allows a lead agent to break down large-scale objectives and delegate specific sub-tasks to specialist agents. These specialists can work in parallel on shared filesystems while maintaining persistent event logs, which provides developers with full transparency into the delegation process. Coupled with new webhook notifications, these tools are designed to streamline complex operations that once required significant human oversight.

These updates represent a major leap toward truly autonomous agentic systems that can learn and adapt in real-time. Whether it is through Dreaming in Managed Agents or more complex orchestration strategies, these tools are helping teams—such as those at Harvey and Spiral—significantly boost completion rates and maintain high editorial standards at scale.

New in Claude Managed Agents. Start building your next autonomous workflow today by visiting the Claude Platform.

Summary

• Dreaming creates a feedback loop for agents to refine their memory and performance across different sessions.
• Outcomes allow for objective, rubric-based grading that drives autonomous self-correction.
• Multiagent orchestration enables scaling by delegating work to specialized agents running in parallel.

Source: New in Claude Managed Agents: dreaming, outcomes, and multiagent orchestration | Claude