Descriptions:
Isabella He, a member of technical staff on Anthropic’s Applied AI team, leads this hands-on workshop introducing Claude Managed Agents—Anthropic’s production-ready infrastructure layer for building and deploying agents at scale. The session traces the evolution from the raw Messages API through the Agent SDK to the current managed harness, explaining why each layer became necessary as model capabilities and agentic task complexity grew.
The bulk of the talk is a live coding walkthrough building an incident response agent. He covers the core primitives: defining agent and environment configurations, setting up sandboxed networking with allow-lists, attaching files via the Files API for log analysis, and binding agents and environments into sessions. A key architectural point is that Claude Managed Agents operates in discrete events—tool calls, user messages, agent responses—rather than raw token streams, enabling more structured observability in production. The session also covers MCP tunnels for private-network MCP server deployments and context engineering strategies for maximizing agent effectiveness.
Viewers will come away understanding how the managed harness handles scaling, compaction, sandboxing, and tool runtime automatically, freeing developers to focus on task configuration and domain-specific tool logic. The talk also previews upcoming managed agents features including “dreaming,” a self-improvement and persistent memory capability being covered in a follow-on session.
📺 Source: Claude · Published May 26, 2026
🏷️ Format: Tutorial Demo
