Claude Managed Agents enter public beta with pre‑tuned agents and infra connectors

Anthropic is trying to productize a thing a lot of companies have been rebuilding badly — the scaffolding around AI agents. Claude Managed Agents, now in public beta, is Anthropic’s hosted service for long-horizon agent work on the Claude Platform. The pitch is simple: let Anthropic run the agent loop, the execution environment, and the state management, so customers don’t have to wire up their own fragile backend. (anthropic.com) ### What actually launched? Claude Managed Agents is not a new model. It’s a managed service that runs agents on a customer’s behalf. Anthropic describes it as a hosted layer for long-running work — the kind of task that takes many steps, uses tools, edits files, and can stretch across multiple context windows instead of finishing in one chat turn. (anthropic.com)e the hard part of agents is no longer just “make the model smart.” It’s keeping the surrounding system from going stale. Anthropic says harnesses — the code that loops the model, routes tool calls, and manages memory — bake in assumptions about what Claude can and can’t do. As models improve, those assumptions stop being useful and start becoming dea(anthropic.com)k that helped Sonnet 4.5 became unnecessary on Opus 4.5. (anthropic.com) ### What’s a harness, exactly? Basically, it’s the agent’s operating loop. The harness calls Claude, decides what to do with Claude’s tool requests, and keeps the task moving. That sounds boring, but it’s where a lot of real-world agent systems break. If your harness assumes the model needs certain workarounds forever, you end up freezing last quarter’s model limitations into this quarter’s pr(anthropic.com)om the “hands.” (anthropic.com) ### What did Anthropic build instead? Anthropic says it “virtualized” the main pieces of an agent into three parts: the session, the harness, and the sandbox. The session is the append-only log of what happened. The harness is the control loop. The sandbox is the execution environment where Claude can run code and edit files. The point is stability at the interface level — Anthropic can change(anthropic.com)ps. (anthropic.com) ### Why does that matter to enterprises? Because most companies do not want to become experts in autoscaling agent containers, recovering failed runs, and securing tool execution just to automate a workflow. Anthropic’s engineering writeup says its first approach bundled everything into one container. That made the system easy to start, but brittle — if the container died, the session died wit(anthropic.com)problem. (anthropic.com) ### Where do the guardrails come in? The safety story sits in the sandbox and tool layers. Anthropic has already been moving Claude Code toward tighter filesystem and network isolation, and it says sandboxing reduces risky autonomy tradeoffs by putting hard boundaries around what the agent can touch. That matters more when agents are running longer and acting more independently. Managed Agents(anthropic.com)rower rails. (anthropic.com) ### Is this different from Anthropic’s older agent advice? Yes — and that’s the interesting part. Back in late 2024, Anthropic’s public guidance leaned toward simple, composable agent patterns over heavy frameworks. Managed Agents doesn’t really reverse that view, but it does say the company now thinks the infrastructure layer itself should be hosted and abstracted for customers. In ot(anthropic.com 1)(anthropic.com 2) ### What’s the bottom line? This is Anthropic moving up the stack. Not just selling a model, but selling the runtime around the model. If that works, enterprises get to buy “agent operations” as a service instead of assembling it from SDKs, containers, and hope. (anthropic.com)