Anthropic ships Claude managed agents

AI agents are starting to look less like prompts and more like infrastructure. That is the real story here. Anthropic has now put Claude Managed Agents front and center in its developer docs, with a quickstart that walks builders through creating an agent, setting up an environment, starting a session, and streaming the results back into an app. ### What is a managed agent, exactly? Basically, it is Anthropic saying: stop wiring the whole agent loop yourself. The managed version gives Claude a hosted runtime where it can use tools, read files, run commands, browse, and keep working across stateful sessions with persistent event history. Anthropic is explicitly positioning this against the plain Messages API, which is still there for teams that want full control. (platform.claude.com) ### Why does that matter? Because the annoying part of agents is rarely the model. It is the scaffolding around the model — tool execution, runtime isolation, state management, streaming, retries, and all the glue code that makes a long task survive contact with production. Anthropic’s engineering write-up makes that point pretty directly: model-facing “harnesses” go stale as models improve, so the better abstraction is to keep the interface stable while the execution layer evolves underneath. (platform.claude.com) ### What changed this week? The docs are now live as a proper product surface, not just an idea. The overview page describes Managed Agents as a pre-built, configurable harness for long-running and asynchronous work, and the quickstart turns that into a concrete recipe. Anthropic also tied Managed Agents into the main API overview, which now lists direct programmatic access to both Claude models and Claude Managed Agents. (anthropic.com) ### Why pair this with Microsoft 365? Because enterprise work still happens in spreadsheets, decks, docs, and inboxes. Anthropic’s new finance push says Claude can now work directly in Excel, PowerPoint, Word, and Outlook through Microsoft 365 add-ins, with Outlook still described as “coming soon” in one announcement. The pitch is continuity — start in a model, end in a presentation, and keep the same working context across apps. (platform.claude.com) ### What are the finance agents for? They are prebuilt workflows aimed at the kind of repetitive, high-stakes tasks banks and investment teams already do. Anthropic says there are 10 of them, covering things like pitch building, meeting prep, earnings review, model building, market research, KYC, onboarding, reconciliation, diligence, and credit memos. That matters because it turns “agent” from a vague capability into something a finance team can trial tomorrow. (anthropic.com) ### Is this the same as chatbots getting smarter? Not really. A chatbot answers turns. A managed agent owns a task. That sounds subtle, but it changes the product shape completely. Once the system has sessions, tools, file access, and background execution, you need controls that look a lot more like cloud software — quotas, audit trails, permissions, and safe runtimes. Anthropic’s docs and product pages are all moving in that direction. (anthropic.com) ### So what is Anthropic really shipping? Turns out it is shipping a platform layer. The model is still the brain, but Anthropic is now selling the hands, the memory, and the workbench too. That is a bigger move than a new model release, because it makes Claude harder to swap out and easier to embed inside real enterprise processes. ### Bottom line The headline is not just “Anthropic launched another agent feature.” It is that Claude is becoming a managed execution environment for office and finance work — and that is where the real enterprise moat starts. (platform.claude.com)