Anthropic targets agent control plane

Anthropic spent the past six weeks adding more of the software layer that sits around AI models, not just the models themselves. On April 8, the company launched Claude Managed Agents in public beta, and on May 6 it added dreaming, outcomes, multiagent orchestration and webhooks to that product. Anthropic says those tools handle secure execution, state management, permissions, tracing and tool use on its infrastructure rather than leaving customers to assemble those pieces themselves. VentureBeat argued on May 16 that this is where the next enterprise fight is moving: into the “agent control plane,” the layer that governs how agents call tools, access data and run workflows. ### What is Anthropic actually shipping beyond the Claude model? Claude Managed Agents is a hosted service for “building and deploying cloud-hosted agents at scale,” Anthropic said in its April 8 announcement. The company said developers define tasks, tools and guardrails, while Anthropic provides sandboxed code execution, checkpointing, credential management, scoped permissions and tracing. Anthropic said a built-in orchestration harness decides when to call tools, how to manage context and how to recover from errors. (claude.com) The same announcement said agents can run for hours, preserve progress through disconnections and coordinate with other agents in parallel for more complex jobs. On May 6, Anthropic added another layer on top of that runtime. The company said “dreaming” reviews past sessions to find patterns and improve memory, “outcomes” lets developers define a rubric for success and have a separate grader evaluate results, and multiagent orchestration is now available to developers building with Managed Agents. (claude.com) ### Why does MCP keep showing up in this story? Anthropic open-sourced the Model Context Protocol, or MCP, on Nov. 25, 2024, describing it as “an open standard for connecting AI assistants to the systems where data lives.” The company said MCP was meant to replace one-off integrations with a standard way to connect models to business tools, repositories and development environments. (claude.com) Anthropic’s own product material now describes MCP as the connectivity layer and “skills” as the workflow layer. (claude.com) In a Dec. 19, 2025 post, the company said MCP connects Claude to third-party tools while skills teach Claude how to use those tools inside a team’s workflow. The MCP project’s March 9, 2026 roadmap shows the same shift toward enterprise operations. Lead maintainer David Soria Parra wrote that 2026 priorities include transport scalability, agent communication, governance maturation and enterprise readiness, reflecting production deployments rather than early experiments. (anthropic.com) ### Where does Microsoft fit in? Microsoft used nearly the same language in a March 9 security blog post announcing Agent 365. (claude.com) The company called Agent 365 “the control plane for agents” and said customers need a way to observe, govern and secure agents across an organization, including those built on Microsoft platforms and by ecosystem partners. (blog.modelcontextprotocol.io) Microsoft said Agent 365 reached general availability on May 1, 2026, as part of its Frontier Suite. The product includes an agent registry, observability features and governance controls, according to the company’s blog post. VentureBeat reported on May 16 that Microsoft remains the enterprise default in its Q1 2026 orchestration tracker and cited Microsoft’s David Weston on the need for a unified control layer. (microsoft.com) That framing places Anthropic’s recent product work in a market where Microsoft already has distribution and enterprise security tooling. ### What does “control plane” mean in practice for enterprises? (microsoft.com) Anthropic’s April 8 post listed the operational pieces directly: secure sandboxing, authentication, tool execution, long-running sessions, scoped permissions and execution tracing. Those are the mechanics that determine whether an agent can do work reliably inside a company, not just answer a prompt. Microsoft’s March 9 post described the same category in governance terms. (venturebeat.com) The company said CIOs and CISOs are asking how to track agents, monitor behavior, control access, prevent data leakage and govern deployments across the enterprise. Anthropic’s Dec. 19 post on skills and MCP put the workflow angle more plainly: tool access alone is not enough if the model does not know which system to query, what to extract and how to format the result. (claude.com) That is the layer where orchestration, state and policy matter. ### What happens next? May 6 is the latest concrete Anthropic milestone in this push. The company said dreaming is in research preview, while outcomes, multiagent orchestration and webhooks are being opened to developers using Managed Agents on the Claude Platform. (microsoft.com) March 9 is the latest roadmap marker for MCP itself. The project said working groups will drive 2026 deliverables in transport, governance and enterprise readiness, while Microsoft’s Agent 365 has already moved to general availability as of May 1. (claude.com) (blog.modelcontextprotocol.io) (claude.com)