Anthropic launches Agents beta

Anthropic has spent more than a year telling developers not to overcomplicate artificial intelligence agents, and now it is selling the part most teams kept rebuilding anyway: the runtime that keeps an agent alive after the demo. On April 8, 2026, Anthropic opened Claude Managed Agents in public beta as a hosted service on the Claude Platform. (anthropic.com) An agent is just a language model with tools and enough time to keep working, like giving a smart intern a browser, a terminal window, and a checklist instead of asking for one answer in one shot. Anthropic’s own 2024 guidance split “workflows,” which follow fixed code paths, from “agents,” which choose their own next steps with tools over longer stretches of time. (anthropic.com) That distinction sounds simple until a team tries to run one in production for an hour instead of a minute. Anthropic’s docs say Managed Agents is for jobs that run for minutes or hours, keep state across multiple interactions, and need secure cloud containers instead of a homemade loop around an application programming interface call. (platform.claude.com) The pitch is that developers stop wiring up the plumbing and start defining the job. Anthropic’s overview says the managed setup replaces the agent loop, tool execution layer, and runtime, while giving Claude built-in access to file operations, shell commands, web search, web fetch, and Model Context Protocol servers for outside tools. (platform.claude.com) The company got here by discovering that the “one big container” approach breaks the moment the work gets messy. In Anthropic’s engineering writeup, the first design put the session, the harness, and the sandbox in one container, and a container failure could wipe out the session entirely. (anthropic.com) So Anthropic split the system into three parts that can survive separately. Its engineering post describes a session as an append-only log, a harness as the loop that calls Claude and routes tool calls, and a sandbox as the execution environment where Claude runs code and edits files. (anthropic.com) That is the same trick operating systems used decades ago when computers stopped being single-purpose machines. Anthropic explicitly compares Managed Agents to stable abstractions like processes and files, where the interface stays the same even as the hardware underneath keeps changing. (anthropic.com) The timing also fits Anthropic’s product line over the last year. When it launched Claude 4 in May 2025, the company leaned hard into long-running agent work, saying Claude Opus 4 could sustain complex tasks for several hours and adding tools like code execution, Model Context Protocol connectors, and prompt caching to support that style of use. (anthropic.com) Anthropic then spent late 2025 reducing one of the biggest hidden costs in agent systems: tool overhead. In November 2025, it said just five connected servers could consume about 55,000 tokens before a conversation even started, which is why it added on-demand tool search and programmatic tool calling. (anthropic.com) Managed Agents turns those separate pieces into a hosted product with its own application programming interface shape. Anthropic’s quickstart shows new endpoints for agents, environments, and sessions, and the beta currently requires the `managed-agents-2026-04-01` header on requests. (platform.claude.com) That makes this less like a new chatbot feature and more like Anthropic trying to become the cloud layer for autonomous work. The company is betting that the hard part of agents is no longer getting a model to sound smart for 30 seconds, but keeping it useful, recoverable, and secure for the next three hours. (platform.claude.com)