Claude Managed Agents and routines

Anthropic is pushing Claude deeper into enterprise work with a hosted agent service and new coding automations, even as users publicly question Claude’s recent reliability. (anthropic.com) (venturebeat.com) On April 9, Anthropic published an engineering post introducing Claude Managed Agents, a hosted service in the Claude Platform for “long-horizon” jobs that run on Anthropic’s infrastructure instead of a customer’s own agent stack. Anthropic said the service is built around three abstractions — session, harness, and sandbox — so the interface can stay stable while the underlying implementation changes. (anthropic.com) Claude Code also got a new “routines” feature on April 14 that lets users save a prompt, repositories, and connectors as a repeatable automation that runs automatically on Anthropic’s cloud. The same update added a redesigned desktop app with a sidebar for multiple sessions and drag-and-drop panes for parallel work. (siliconangle.com) (macrumors.com) An AI agent is software that can plan steps, call tools, and keep working across many turns instead of answering one prompt at a time. Anthropic has been steering Claude toward that model for months: when it launched Claude Opus 4.6 on February 5, it pitched longer-running coding, a one million token context window in beta, and “effort controls” that let developers trade off speed, cost, and reasoning depth. (anthropic.com 1) (anthropic.com 2) The new products shift more of that setup work from customer teams to Anthropic. In its Managed Agents post, Anthropic argued that custom agent “harnesses” can become stale as models improve, so a hosted service can swap in new infrastructure without forcing enterprises to rebuild their own orchestration layer. (anthropic.com) That pitch is landing as some of Claude’s most active users say the product has become less dependable. VentureBeat reported complaints across GitHub, X, and Reddit that Claude Opus 4.6 and Claude Code now stop early more often, reason less deeply, and waste more tokens, while Anthropic employees denied the company was downgrading models to manage demand. (venturebeat.com) One widely shared complaint came from Stella Laurenzo, identified by VentureBeat as an Advanced Micro Devices artificial intelligence executive, who posted a GitHub analysis of 6,852 Claude Code session files, 17,871 thinking blocks, and 234,760 tool calls. Her post argued that reasoning depth fell after February and that Claude Code had become harder to trust for complex engineering work. (venturebeat.com) Anthropic has acknowledged some product changes around usage and defaults, though not the broader claim that Claude was intentionally “nerfed.” The Register reported that Anthropic cut Claude Code prompt-cache time-to-live for many requests from one hour to five minutes around March 7, and Claude Code creator Boris Cherny said the company was also investigating a 400,000-token default context window with an option for one million tokens. (theregister.com) The company’s own status page shows a rough stretch in the days around the launch. Anthropic logged elevated model-request errors on April 10, an email login outage on April 11, a Claude.ai and Claude Code login incident on April 13, and degraded performance on usage and analytics admin application programming interface endpoints on April 14. (anthropic.statuspage.io) Anthropic is also telling customers that Claude Code is already being used at scale. Its product page says Stripe deployed Claude Code to 1,370 engineers, and Ramp used it to cut incident-investigation time by 80 percent. (anthropic.com) That leaves Anthropic trying to sell convenience and control at the same time: fewer moving parts for customers, but more dependence on Anthropic’s own systems and tuning choices. If Managed Agents and routines work as advertised, they reduce engineering overhead; if performance complaints keep spreading, they give enterprise buyers one more reason to ask how much of the stack they want to hand over. (anthropic.com) (fortune.com)