Anthropic launches managed agents
Anthropic opened Claude Managed Agents to public beta, pitching APIs and orchestration tools that let organisations deploy and sandbox multi‑agent workflows instead of just single chatbots. The move signals a shift from model-only competition to selling production-ready agent platforms with orchestration, governance and multi-agent support. That matters because enterprises now prize deployability and control as much as raw capability. (x.com)
Most companies can get a chatbot demo working in a day. The part that takes months is everything around it: the locked-down computer it runs in, the memory of what happened last hour, the permissions for tools, and the recovery plan when it breaks. (anthropic.com) Anthropic’s new product is aimed at that mess, not at chat itself. On April 8, 2026, it opened Claude Managed Agents in public beta as a hosted service that runs long-horizon agents on Anthropic’s own infrastructure. (claude.com) The basic pitch is simple: the customer defines the job, the tools, and the rules, and Anthropic runs the agent. Anthropic says that setup can cut the path from prototype to production from months to days. (claude.com) Anthropic’s docs break the system into three pieces. An “agent” is the model plus its prompt and tools, an “environment” is the container it runs in, and a “session” is the live instance doing one task and producing outputs. (platform.claude.com) That split matters because AI agents fail in boring ways. Anthropic says old agent “harnesses,” meaning the control loop wrapped around the model, can go stale as models improve, so it built Managed Agents around interfaces it hopes will survive model changes underneath. (anthropic.com) One of those interfaces is the sandbox, which is basically a sealed workroom for the model. Anthropic says the built-in toolset can include bash, file operations, and web search, so the agent can do real work without being dropped straight onto a company’s production systems. (platform.claude.com) Another piece is persistence. Anthropic says sessions can keep running for hours and keep their outputs even if the client disconnects, which is the difference between a chatbot that answers one prompt and an agent that can finish a report, a coding task, or a spreadsheet job while you are away. (reworked.co) Anthropic is also pushing beyond one agent doing one job. Its launch materials say Managed Agents can coordinate multiple specialist agents in parallel, though that multi-agent feature is still only in research preview rather than the main public beta. (reworked.co) The company is pricing this like infrastructure, not like a consumer app. Reworked and other launch coverage report standard Claude token charges plus $0.08 per active session-hour, which makes the target customer look less like an individual developer and more like a team paying to skip backend engineering work. (reworked.co) That fits the rest of Anthropic’s business this year. In March 2026 it committed an initial $100 million to its Claude Partner Network, and in late 2025 and early 2026 it announced enterprise pushes with Accenture, Snowflake, Infosys, Salesforce, Deloitte, and Cognizant around getting AI from pilot to production. (anthropic.com 1) (anthropic.com 2) (anthropic.com 3) So this launch is less about “here is a smarter model” than “here is the operating layer for AI workers.” Anthropic is betting that the next fight in artificial intelligence is not just whose model scores highest, but whose platform can safely run agents inside a real company on Wednesday morning. (anthropic.com)
