Anthropic bundles Managed Agents

Anthropic is trying to sell companies the part they usually fail at: not the demo, but the handoff from a clever prototype to a system that can keep running after midnight without an engineer babysitting it. On April 8, Anthropic introduced Claude Managed Agents as a hosted service for “long-horizon” work, meaning jobs that unfold over many steps instead of one chat reply. (anthropic.com) That “managed” part is the pitch. Instead of a customer wiring together its own loops, memory, tools, and execution environment, Anthropic says it will run the agent on the Claude Platform through a small set of stable interfaces. (anthropic.com) Anthropic’s engineers describe the old problem in blunt terms: the scaffolding around an agent keeps breaking as the model gets better. A workaround built for Claude Sonnet 4.5, including “context resets” to stop premature wrap-ups near the context limit, became unnecessary when the company switched the same harness to Claude Opus 4.5. (anthropic.com) So Anthropic split the system into three parts. A session stores the running log, a harness is the loop that calls Claude and routes tool use, and a sandbox is the place where Claude can run code and edit files. (anthropic.com) The analogy Anthropic uses comes from operating systems. Old computers changed from disk packs to solid-state drives, but the `read` command stayed the same, and Anthropic wants agent infrastructure to work the same way even as the models and tool logic underneath keep changing. (anthropic.com) This launch also fits a pricing pattern Anthropic started pushing last year. In August 2025, it said Team and Enterprise customers could buy seats for the Claude app, then turn on extra usage at standard application programming interface rates with spend caps, which is closer to cloud billing than classic flat-fee software subscriptions. (anthropic.com) That matters because agents burn money in bursts. A company might want 500 employees on predictable seat licenses for everyday chat, but only a smaller group running coding agents, research agents, or long jobs that consume large volumes of tokens and need metered billing. (anthropic.com; anthropic.com) Anthropic has been laying the groundwork for this for months. Its December 2024 guide on building effective agents argued that the best systems usually use simple, composable patterns instead of giant frameworks, and its November 2025 post on long-running agents focused on the problem of work stretching across hours or days with no memory between sessions. (anthropic.com; anthropic.com) The company has also been turning those ideas into products around work, not just models. Anthropic’s site now lists Claude Code for Enterprise, research tooling, and business integrations alongside the core Claude app, which shows how the company is packaging the model inside workflows people can actually buy. (anthropic.com; anthropic.com) Wired reported that Anthropic’s annualized recurring revenue has climbed to about $30 billion, which is an unusually large number for a company still early in selling agent software to enterprises. Wired tied that growth to products that wrap models inside business workflows rather than selling raw intelligence alone. (wired.com) Anthropic also raised $30 billion in Series G funding on February 12, 2026, at a $380 billion post-money valuation, giving it the capital to subsidize infrastructure-heavy enterprise products while it races to lock in big customers. Managed Agents looks like one more step in that race: less “here is a model,” more “here is the employee-shaped software around it.” (anthropic.com; anthropic.com)