Descriptions:
Michael and Harrison, both members of technical staff at Anthropic, deliver a keynote-style deep dive into Claude Managed Agents — the company’s production platform for building, deploying, and scaling AI agents. The talk opens with a capability retrospective from Claude 3 through Opus 4.7, arguing that the binding constraint on agent performance has shifted from model intelligence to surrounding infrastructure.
The presentation systematically covers the platform’s core primitives: multi-agent orchestration for delegating work across specialized subagent threads; Outcomes, a goal-rubric loop that iterates until completion criteria are satisfied; Memory, a persistent file-based learning system that improves performance across sessions; and Dreaming, a research-preview feature that runs globally across thousands of session transcripts to curate and refine agent memory beyond what any single agent could achieve locally. Two major feature announcements anchor the talk: self-hosted sandboxes — enabling teams to bring their own compute via Cloudflare, Daytona, Modal, or Vercel rather than running on Anthropic infrastructure — and MCP Tunnels, which expose private internal MCP servers to Claude without any public internet exposure.
Partner testimonials and a getting-started walkthrough close out the session, making this a strong reference point for any engineering team evaluating Claude Managed Agents as a foundation for production agentic systems.
📺 Source: Claude · Published May 21, 2026
🏷️ Format: Keynote Launch
