Anthropic opens Managed Agents beta
Anthropic has put Claude Managed Agents into public beta, letting companies scale AI agents from prototype to production with a managed platform. The move marks another push by an AI vendor to make agent architectures easier to deploy at scale, which could accelerate how firms automate routine white-collar tasks. The public beta framing suggests Anthropic is aiming to move more customers from experiments into live workflows. (x.com/claudeai/status/2041927687460024721)
Anthropic just moved one of the messiest parts of artificial intelligence from custom code into a hosted product: the loop that keeps an agent running for a long task instead of answering once and stopping. On April 9, 2026, the company said Claude Managed Agents is now in public beta. (anthropic.com) An artificial intelligence agent is a model plus tools plus memory, which is closer to hiring a junior operator than sending a single prompt. Anthropic says Managed Agents is a hosted service for “long-horizon” work, meaning jobs that unfold over many steps instead of one reply. (anthropic.com) That sounds simple until the agent has to remember what already happened, call software tools, run code, and survive crashes without losing its place. Anthropic says it built the service around three pieces: a session, a harness, and a sandbox. (anthropic.com) The session is the running log, like a flight recorder that stores every event in order. Anthropic describes it as an append-only record, so a failed run can resume from the last saved step instead of starting over. (anthropic.com) (aihola.com) The harness is the traffic cop that decides when to call Claude and when to send Claude’s tool requests to outside systems. Anthropic’s pitch is that this middle layer should be replaceable, because the tricks developers use today can become obsolete when models improve. (anthropic.com) Anthropic gives one concrete example of that problem from its own earlier work. It says Claude Sonnet 4.5 needed “context resets” to avoid wrapping up too early near a context limit, but Claude Opus 4.5 no longer showed that behavior, which turned the workaround into dead weight. (anthropic.com) The sandbox is the workroom where the agent can run code and edit files without touching the rest of the system directly. Anthropic says its first design put the session, harness, and sandbox in one container, and that created a “pet” server that was painful to lose when it failed. (anthropic.com) (aihola.com) That architecture problem matters because companies do not want a clever demo that dies halfway through a spreadsheet job or software task. Anthropic says separating state from execution lets a new harness pick up the same session after a crash, while the sandbox becomes just another tool the system can call. (anthropic.com) (aihola.com) This launch also fits a longer Anthropic push to make agents less brittle and less expensive to run. In November 2025, the company added tool search and programmatic tool calling, and said one five-server setup could burn about 55,000 tokens before a user even asked a question. (anthropic.com) The model side has been moving in the same direction. When Anthropic introduced Claude 4 on May 22, 2025, it said Opus 4 could sustain long-running tasks for hours and added application programming interface features like code execution, Model Context Protocol connectors, and prompt caching to support stronger agents. (anthropic.com) Put together, Managed Agents is Anthropic trying to sell not just the brain but the plumbing. The bet is that more companies will pay for a managed runtime that handles sessions, orchestration, and sandboxes if that means moving from prototype agents to production systems faster. (anthropic.com) (siliconangle.com)
