Anthropic launches Managed Agents

Anthropic is no longer just selling a model you call with an application programming interface. On April 8, 2026, it put Claude Managed Agents into public beta as a hosted service that runs long-horizon agents for you on the Claude Platform. (anthropic.com) That changes one boring but expensive part of building agents: the scaffolding around the model. Anthropic says developers can define the task, tools, and guardrails, while Anthropic handles the runtime pieces that usually turn a demo into months of backend work. (anthropic.com) An agent is just a language model that can keep working after one reply. Instead of answering one prompt and stopping, it can call tools, write code, edit files, and keep a running record of what happened across a session. (anthropic.com) The hard part is not only the model. The hard part is the loop around it, which Anthropic calls a harness: the software that decides when to call Claude again, when to use a tool, where to store state, and how to recover when something breaks. (anthropic.com) Anthropic’s own engineers have been writing about this problem since 2025 because long-running agents hit practical limits fast. In a November 2025 post, the company said agents working across many context windows still needed special harness design for jobs that run for hours or days. (anthropic.com) Managed Agents packages that harness into a service with three parts. Anthropic says it virtualized the session, which is the append-only log of everything that happened, the harness, which is the tool-calling loop, and the sandbox, which is the execution environment where Claude can run code and edit files. (anthropic.com) The company’s engineering post uses an operating system analogy. Old operating systems hid changing hardware behind stable abstractions like files and processes, and Anthropic says Managed Agents tries to do the same thing for agent infrastructure as models improve underneath. (anthropic.com) That model-improvement point is not theoretical in Anthropic’s telling. The company says a fix it built for Claude Sonnet 4.5, adding context resets to stop tasks from ending early near the context limit, became unnecessary on Claude Opus 4.5, which meant the old harness logic had already gone stale. (anthropic.com) Anthropic also says it moved away from putting everything in one container because that made each running agent a fragile “pet” server. If that container failed, the session was lost, so the company redesigned the system around more interchangeable components. (anthropic.com) The business signal is that Anthropic wants to own more of the stack than tokens. Its March 2026 State of AI Agents report, based on a survey of more than 500 technical leaders in the United States, described the market as shifting from single-function pilots toward cross-functional deployment, and Managed Agents is a product built for that phase. (resources.anthropic.com) Anthropic has been pushing Claude as a model for “coding, agents, and professional work,” not just chat. Its current product pages for Claude Sonnet 4.6 and Claude Opus 4.6 both position the models around agentic workflows, so Managed Agents is the infrastructure layer that matches that pitch. (anthropic.com 1) (anthropic.com 2) If this works, companies stop treating agents like clever prompts and start treating them like workloads. Anthropic’s announcement turns the question from “which model should answer this?” into “who runs the session, sandbox, recovery, and lifecycle when 1,000 agents are working at once?” (anthropic.com)