Anthropic: Managed Agents

Most companies can get a language model to answer a prompt. The hard part starts when that model has to keep working for 20 minutes, use tools, survive a crash, and pick up where it left off without losing the thread. (platform.claude.com) That missing layer is called the runtime: the part that stores state, runs tools, manages permissions, and keeps an agent alive between steps. Anthropic’s news on April 8, 2026 was that it wants to host that layer itself with Claude Managed Agents, now in public beta. (anthropic.com) (wired.com) Anthropic’s pitch is simple: instead of writing your own agent loop, you define the job and Anthropic runs the agent in its cloud. The official docs say the managed environment can read files, run commands, browse the web, and execute code securely. (platform.claude.com) The old way looked a lot like building your own mini operating system around a model. Anthropic says production agents usually need sandboxed code execution, checkpointing, credential management, scoped permissions, and tracing before a customer sees anything useful. (anthropic.com) (thenewstack.io) Checkpointing is the key idea here. It works like a video game save point: if the agent fails halfway through a long task, the system can resume from the saved state instead of starting over from the first step. (wired.com) (platform.claude.com) Anthropic is pairing that hosted runtime with something it calls Agent Skills. A skill is a reusable package of instructions, metadata, and optional scripts that Claude can automatically pull in when a task matches that skill. (platform.claude.com) That turns agent building into more of a composition problem than a plumbing problem. Instead of hand-coding every workflow from scratch, a team can combine a managed runtime with prebuilt or custom skills for tasks like document work, coding, or internal company processes. (platform.claude.com) (github.com) Anthropic’s engineering post describes this as separating the “brain” from the “hands.” The model decides what to do, while the hosted system handles the messy mechanics of actually doing it across tools, files, and long-running sessions. (anthropic.com) That matters because a lot of enterprise agent work has been custom glue code. Wired reported that Anthropic is trying to lower the barrier for businesses that want agents without staffing a team just to manage memory, recovery, and orchestration. (wired.com) The broader bet is that agents become something like serverless computing did for web apps: you stop managing the machine and start describing the job. If Anthropic can make long-running agents feel like calling an application programming interface instead of babysitting infrastructure, it moves the competition from raw model quality to who can offer the easiest production system. (thenewstack.io) (platform.claude.com)