Overview of New Features in Claude Managed Agents

This document outlines the technical enhancements to Claude Managed Agents, introducing features that improve agent autonomy, task accuracy, and system efficiency. The primary advancement involves moving toward complex, multi-agent workflows with automated self-improvement and memory optimization. Key Concepts: The update introduces Multi-Agent Orchestration, allowing lead agents to manage sub-agents with shared filesystems. The 'Outcomes Loop' provides a rubric-driven quality control mechanism, while 'Dreaming' facilitates background memory curation. Webhooks are implemented to improve communication efficiency. Procedures: 1. Multi-Agent Orchestration: A lead agent delegates tasks to specialized sub-agents; while sub-agents hold isolated context windows, they operate within a shared container and filesystem. 2. Outcomes Loop: A grader sub-agent evaluates generated output against a predefined rubric and provides iterative feedback until standards are met. 3. Dreaming: The agent processes background reflections, allowing it to edit, add, or curate memory based on identified patterns from past sessions. 4. Webhooks: Developers can configure notifications for task completion, removing the requirement for active status polling. Entities and Tools: The update pertains specifically to Claude Managed Agents, as reported by Wes Roth. Technical terms include Multi-Agent Orchestration, Outcomes Loop, Dreaming, and Webhooks. Facts: These features enable agents to perform self-directed improvement, manage persistent memory through reflection, and communicate progress asynchronously, significantly reducing overhead associated with manual task monitoring and constant polling.