Descriptions:
Anthropic’s official Claude channel has released a product walkthrough for Claude Managed Agents, a new suite of APIs designed to let developers build and deploy AI agents at scale inside isolated, configurable sandbox environments. The demo shows agents running in containers with full file system access, bash execution, and web search — all provisioned from environment definitions that specify pre-installed packages, network controls, and mounted repositories before a session begins.
Three concrete scenarios are demonstrated. First, a Kanban board integration where dragging a ticket to “in progress” automatically fires a Claude session: the agent runs a Lighthouse audit, compresses images, inlines CSS, and iterates until a rubric score of 96 is reached — with a separate grader evaluating output against defined criteria in its own context window. Second, a weekly SaaS pricing tracker that searches current pricing pages, runs a Python cost analysis in the sandbox, writes an Excel summary, and posts a report to Slack and Asana via MCP servers. Third, a multi-agent incident response system where a coordinator delegates to three specialist sub-agents running in parallel context windows on a shared file system, with a human-approval permissions gate before any message is posted publicly.
Memory is a first-class feature throughout: agents check a persistent memory store before starting each run and update it afterward, so recurring workflows begin with accumulated context. Managed Agents targets developers who want a fully managed, stateful agent platform without building session management, sandboxing, or orchestration infrastructure themselves.
📺 Source: Claude · Published April 09, 2026
🏷️ Format: Keynote Launch
