Agents in the Cloud: A Safer Setup When Everyone's a Developer

Not long ago, AI assistants simply answered your questions in a browser. The early versions had no access to your files or the internet. Then they were connected to external knowledge and your company data and could call APIs and databases to take real actions. 

Today’s coding agents are even more capable. They plan, break objectives into multiple steps, invoke multiple tools, fix their own errors, and hand work off to other agents to autonomously complete complex tasks. 

At the same time, code-capable agents like Claude Cowork and OpenAI Codex have moved from developer tools to software everyone uses every day. But the underlying infrastructure they run on has remained largely the same. We think the next step is clear: these agents should run in managed cloud environments, not on the endpoint.

A New Frontier of Risk

This shift in agent capabilities and who’s using them brings serious security implications.

The mistakes of early AI models were passive: a model could make up a historical event or give bad advice, but it had no way to act on that. Now, anyone can fire off a simple prompt that launches multiple agents to generate and execute code that can surface sensitive data or cause serious harm. Put simply, the risk of an employee causing a security incident has shot up.

Remember that AI companies built agentic coding tools for developers, who are much more aware of security concerns and understand the impact of the code agents write. The finance and HR employees now running the same tools don’t bring that same background knowledge. Agents mean they can easily leak a credential, publish internal data to a public location, or reach a data environment they were never cleared to touch — all without even realizing it.

The Case For Running Agents in the Cloud

To understand the value of cloud-native agents, take sandboxes. The cloud extends agents sandboxes and containers to all employees, not just technical teams who already know how to set these up locally. This makes it easy to centrally control agent access so employees only access the data they need — not the finance director’s salary spreadsheet or a developer's leftover master key.

Network control is also easier in the cloud. Egress is hard to police on an endpoint, where a code-capable agent can often find a way around the rules. In the cloud, egress control is much harder to bypass. Version control is simpler, too. Many organizations still struggle to keep endpoint tools current, but the cloud makes that as easy as the push of a button.

You also inherit the benefits of the top-tier security teams large cloud providers employ. They have the top talent and big budgets needed to keep a constant eye on the newest threats. They also offer strong identity and network controls. These just aren’t things smaller in-house teams can match.

Beyond that, cloud execution is faster to roll out: you can enable your whole staff at once, with little setup required to create agents doing real work for non-technical employees. 

The Tradeoffs

There are a few advantages to running agents locally. Local network control still has real value, and a VPN gives you tighter identity control.

Running agents in the cloud can also pose new problems. The most pressing issue for many companies: agents use a lot of compute, and running them around the clock can get expensive quickly. At enterprise scale, you need strong controls to prevent bills from spiraling out of control in a matter of days.

Second, no sandbox is without vulnerabilities, and container escape remains a real threat. Third, the cloud simply isn’t an option for select organizations that can’t send regulated data, customer records, or sensitive source code to a third-party cloud due to privacy and regulatory requirements.

For most organizations, though, these risks don’t outweigh the major benefits. 

A Market Shift is Already Happening

There are already signs that agentic AI is shifting to the cloud. 

Think about what local execution actually needs: the file system and the network. Both have already moved. Inference runs in the cloud, and most of what an agent touches lives in SaaS and remote services. So at this point, a locally running agent really does most of its work in the cloud. 

The frontier labs have already jumped on board. Anthropic has cloud-hosted Claude Code, where agents run in isolated cloud sandboxes instead of on local machines. OpenAI’s Codex, which launched a few months after Claude Code, has lived in the cloud since the early days.

The advantages are real: Sessions keep running after the user disconnects, repositories no longer need to be cloned locally, and organizations can set network access and execution policy centrally. 

Where This Goes

Agents are becoming standard for every employee, not just engineers, and the architecture needs to modernize accordingly. Running them on individual laptops adds security, operational, and governance problems that a managed cloud environment solves by default. Agents need to be plug and play, batteries included, and the cloud is a much better model for that.