Claude Managed Agents add modular 'Agent Skills' and three new features

Anthropic is pushing Claude further away from “chatbot with tools” and closer to “agent platform you can actually run at work.” The new piece is Agent Skills — modular bundles of instructions, metadata, and optional scripts or templates that Claude Managed Agents can pull in automatically when a task calls for them. The bigger point is control. Companies want long-running agents, but they also want those agents to behave in repeatable, reviewable ways. Anthropic is trying to package that discipline into the product itself. (platform.claude.com) ### What is a Managed Agent? Managed Agents are Anthropic’s hosted agent runtime. Instead of building your own loop for tool calls, sandboxing, file handling, and state management, you define an agent, define an environment, start a session, and let Anthropic run the rest in managed infrastructure. The setup is aimed at jobs that take minutes or hours rather than one prompt(platform.claude.com)at need persistent files, tools, and server-side history. (platform.claude.com) ### Why add Skills on top? Because a general model still forgets the “house style” of a workflow. A company may need Claude to build a board deck, clean a spreadsheet, or follow a compliance checklist the same way every time. Skills are Anthropic’s answer to that gap. They’re filesystem-based resources that give the agent domain-specific workflows, context, and best practic(platform.claude.com)lly, the prompt stops carrying the whole burden. (platform.claude.com) ### What does a Skill actually contain? A Skill can package instructions, metadata, and optional resources like scripts and templates. Anthropic describes them as modular capabilities that Claude uses automatically when relevant. That matters because the model is not just handed a giant wall of guidance at the start. The system can progressively load what it needs from the Sk(platform.claude.com)sting a mega-prompt into chat. (platform.claude.com) ### Are these only Anthropic-made tools? No. Anthropic supports two kinds inside Managed Agents: pre-built Anthropic skills and custom skills uploaded by an organization. The pre-built set covers common document tasks — PowerPoint, Excel, Word, and PDF handling. Custom skills can be versioned, pinned, and attached alongside Anthropic’s own. A session can i(platform.claude.com)ed pattern: compose several narrow capabilities instead of one giant monolith. (platform.claude.com) ### Why does this matter for enterprises? Because the hard part of enterprise agents is not just capability — it’s governance. Anthropic now has dedicated enterprise guidance for reviewing Skills, including checks for code execution, hidden instructions, network calls, MCP references, file-system scope, and hardcoded credentials. The catch is obvious: once you let a Skill bun(platform.claude.com) leaning into that instead of pretending prompt engineering solves it. (platform.claude.com) ### How does this fit Anthropic’s broader product push? Pretty neatly. Anthropic has been framing agents as long-horizon systems that need stable interfaces even as the underlying harness changes. It also keeps expanding Claude into workplace software — including Microsoft 365 add-ins across Excel, PowerPoint, and Word, with Outlook marked as coming soon in(platform.claude.com)pic wants developers and companies to snap onto a managed Claude “brain.” (anthropic.com) ### What’s the bottom line? Anthropic is productizing a simple idea: prompts are too flimsy for serious agent work. Managed Agents give Claude a hosted runtime. Skills give that runtime reusable expertise. Put together, that makes Claude easier to deploy for bounded, longer-running workflows — and easier to audit when something goes wrong. (platform.claude.com)