Anthropic Ships Managed Agents
Anthropic has pushed beyond selling models and started offering managed agents and a Skills system that lets agents call tools and follow instructions, but it warns Skills should only be used from trusted sources because they expand the attack surface. The company frames this as solving the hard operational parts of running agents—monitoring, permissions and reliability—so businesses can produce agent-driven workflows without building the whole infra themselves. That shift matters because the bottleneck for agent products is moving from prompting to orchestration, trust and observability, which changes product and cost trade-offs for builders. (wired.com) (platform.claude.com)
Anthropic is no longer just selling you a model and telling you to wire the rest yourself. On April 8, 2026, it put Claude Managed Agents into public beta as a hosted service that runs long tasks in Anthropic’s own cloud environment. (claude.com) A managed agent is basically “the model plus the scaffolding around it.” Anthropic says customers define the task, tools, and guardrails, while Anthropic handles the agent loop, tool execution, runtime, and production infrastructure. (platform.claude.com) (claude.com) That scaffolding is the part most demos skip. Anthropic’s launch materials list sandboxed code execution, checkpointing, credential management, scoped permissions, and end-to-end tracing as the work companies usually have to build before users ever see a feature. (claude.com) The product is aimed at jobs that do not finish in one chat turn. Anthropic says Managed Agents is best for work that runs for minutes or hours, keeps state across multiple interactions, and needs secure containers, mounted files, network rules, and server-side event history. (platform.claude.com) Inside that setup, Claude can do more than answer text prompts. Anthropic’s documentation says the managed environment lets the agent read files, run shell commands, browse the web, execute code securely, and connect to outside tool providers through Model Context Protocol servers. (platform.claude.com) Anthropic also shipped Skills, which are reusable add-ons that work more like installing a playbook than writing a prompt. A Skill can bundle instructions, metadata, scripts, templates, and reference files so Claude loads the right capability when a task calls for it. (platform.claude.com) Anthropic already includes prebuilt Skills for PowerPoint, Excel, Word, and Portable Document Format files, and developers can make custom ones for company-specific workflows. The company says Claude can create Skills in Claude Code, upload them through the application programming interface, or add them through claude.ai settings. (platform.claude.com) That convenience comes with a security warning. Anthropic’s Skills documentation says Skills can include executable code and filesystem resources, and the company warns developers to use custom Skills only from trusted sources because they expand what the agent can do inside its virtual machine environment. (platform.claude.com) Anthropic’s own engineers describe the bigger goal as separating the “brain” from the “hands.” In an April 8 engineering post, the company said model behavior changes fast enough that hard-coded agent harnesses can become obsolete, so it wants developers to build against stable interfaces while Anthropic keeps swapping out the machinery underneath. (anthropic.com) That is why this launch looks different from a normal model release. The pitch is less “here is a smarter model” and more “here is the operating layer for agents,” which turns reliability, permissions, observability, and recovery from errors into the product instead of leaving them as homework for every developer. (anthropic.com) (claude.com)
