Anthropic launches Claude agents beta

Anthropic is trying to sell companies a shortcut: instead of wiring together servers, schedulers, retries, and security controls for an artificial intelligence agent, they can now rent the whole thing from Anthropic in a public beta that opened on April 8, 2026. The product is called Claude Managed Agents, and Anthropic says it is for cloud-hosted agents that run at scale rather than one-off demos on a laptop. (claude.com) An artificial intelligence agent is just a model that does more than answer one prompt. It can take a goal like “research this company” or “triage these support tickets,” use tools, keep state across steps, and keep going until it finishes or fails. (anthropic.com) That sounds simple in a demo, but the hard part is everything around the model. Anthropic’s engineering team says most agent systems need a “harness,” which is the control layer that decides when the model can call tools, how long it can run, what to retry, and what to do when a step breaks. (anthropic.com) Anthropic’s pitch is that companies should stop rebuilding that harness from scratch every time models improve. In its engineering write-up, the company says those hand-built assumptions go stale fast, so Managed Agents is designed to keep the interface steady while Anthropic changes the orchestration underneath. (anthropic.com) The new service is a suite of application programming interfaces, which are the software hooks developers use to connect one system to another. Anthropic says those interfaces let teams build and deploy cloud-hosted agents faster, with the company handling the runtime layer instead of the customer running its own orchestration stack. (claude.com) Anthropic is putting a concrete number on that promise. In the launch post, the company says Managed Agents can get teams to production “10x faster,” which is a claim about setup speed rather than model intelligence. (claude.com) The pricing model shows who this is for. Reworked, citing Anthropic’s beta rollout, reports standard Claude token charges still apply, and customers also pay $0.08 per active session-hour, which makes the product look more like rented infrastructure for businesses than a flat-fee consumer tool. (reworked.co) Anthropic is also signaling that this is already in real company workflows, not just a developer preview. Coverage of the launch says early users include Notion, Asana, and Rakuten, which are all software-heavy businesses that could plug agents into search, writing, task management, or internal operations. (siliconangle.com) This launch lands in the middle of a broader shift in the artificial intelligence market. Model companies spent 2023 and 2024 selling raw intelligence through chatbots and application programming interfaces, and by 2026 they are increasingly trying to sell the operating system around that intelligence: tools, memory, permissions, monitoring, and billing. (anthropic.com) (9to5mac.com) So the news is not just that Anthropic launched another Claude feature. It is that Anthropic wants Claude to move from “assistant you ask for help” to “service your company deploys,” with Anthropic charging by the hour to keep those agents running in the cloud. (claude.com) (reworked.co)