Anthropic will let your agents sleep on its couch

What’s new

Anthropic has launched Managed Agents, a hosted service that runs long‑running, tool‑enabled AI agents on the company’s infrastructure. An agent, Anthropic explains, bundles a model, a system prompt, tools, checkpoints and skills into a versioned persona you can deploy; the idea is to hand over the messy plumbing — sandboxed execution, credential handling, scoped permissions, tracing — and let Anthropic manage it. Want to run your business on autopilot? For better or worse, that’s the pitch.

How it works

Agents can call tools, read files, run commands, browse the web and execute code inside cloud-hosted secure containers. Developers author sub‑agents with Markdown and YAML to steer behavior — think a frontend specialist versus a general coder — and an orchestration harness decides when to call tools, manage context and recover from errors. Claude Code remains the primary commercial use case, but Anthropic says Managed Agents are meant for a broad set of office tasks; it has been reported that Notion’s product manager praised the service in a YouTube testimonial, describing how the firm uses Managed Agents to ship code and produce websites and presentations.

Why it matters — and why you should worry a little

This is the latest sign that AI is moving from interactive assistants to persistent background workers. That’s exciting: long workflows, repeated tool calls and persistent state can be automated at scale. It’s also unnerving. Unsupervised action means unsupervised spending. Handing keys, credentials and execution rights to an autonomous process — even one running in a “secure” container — raises familiar questions about auditability, scope creep and failure modes. Anthropic says you define guardrails; the rest is on them. Will enterprises sleep easier? Depends on how good the locks are.

Sources: The Register