Cloud-based AI agents are becoming the next major shift in how companies deploy artificial intelligence, with OpenAI, Anthropic, and Cursor all prioritizing this infrastructure change. Instead of running AI models locally on a developer's machine, these agents now execute long-running tasks in remote cloud environments, eliminating setup friction and allowing work to continue without tying up personal hardware.

Why Are Tech Companies Moving AI Agents to the Cloud?

The move away from local AI agents stems from practical frustrations. Developers running multiple AI agents on their personal computers experienced significant slowdowns, with CPU-intensive tasks heating up machines and degrading overall system performance. Peter Steinberger, an AI developer, built Crabbox specifically to address this problem, creating a way to run OpenAI's agents in the cloud instead of locally.

The same solution emerged independently across multiple organizations. At Anthropic, the company has made "Claude Managed Agents" a major focus, with Katelyn Lesse, head of engineering for Claude Platform, explaining that her team spent six months building this hosted service to execute long-running agents across various cloud providers. OpenAI acquired Ona, formerly known as Gitpod, a leader in cloud development environments, specifically to support this transition. The company stated that as AI models become more capable, "its most valuable work is unfolding over hours or days, rather than minutes," requiring persistent cloud environments where agents can access tools and context over extended periods.

What Technical Challenges Do Cloud Agents Face?

Moving agents to the cloud introduces new engineering problems that teams are actively solving. At Cursor, which released Cloud Agents at the end of last year, cofounder Sualeh Asif revealed several unexpected challenges:

• Error Feedback Loops: Local agents can surface warnings and errors directly to humans, who then instruct the model on next steps. Cloud agents lack this immediate feedback mechanism, so Cursor developed a "confess" model where agents regularly report issues to infrastructure teams for environment improvements.
• Node Failure Recovery: Long-running cloud tasks face nontrivial engineering challenges when infrastructure nodes terminate mid-execution, requiring teams to solve how agent execution moves from one node to another without losing progress.
• Asynchronous Execution: Cloud agents operate independently with their own tools and resources, allowing them to run longer and iterate toward production-ready code without constant human intervention.

Cursor recently launched an iOS app that enables software building from anywhere, a product built entirely on top of cloud agents to support long-running tasks.

How Are Companies Adopting Cloud Agents in Practice?

The adoption is accelerating across the industry. OpenAI is actively hiring engineers for its Cloud Agents team, with job postings seeking experienced backend and infrastructure engineers to "design and operate systems for orchestrating agents at scale." The company emphasizes that this is a recent development, with engineers shifting focus from single-device applications to cloud-based orchestration.

At OpenAI's offices, engineers confirmed that their focus is "very much" shifting toward cloud-based agents. The infrastructure requirements are substantial, with teams needing expertise in Python, Rust, distributed systems, and cloud infrastructure to build the platforms that will support this transition.

One concrete example of this shift comes from Anthropic's Claude Tag feature, a Slack integration that allows users to mention Claude and kick off tasks without additional setup. While the Slack integration itself seems simple, the real innovation is the ability to launch AI agents that run in the cloud rather than on local machines, eliminating context-switching and setup friction.

"No additional setup. For Claude Code to work well, it should be connected to internal MCP servers, with the right skills on your local machine. Of course, at larger companies this setup is at least partially automated, but devs often need to do tweaking," explained David Hershey, head of Applied AI at Anthropic, describing why cloud agents represent a meaningful improvement over local alternatives.

David Hershey, Head of Applied AI, Anthropic

What Does This Mean for Software Engineers?

The shift to cloud agents signals a broader transformation in how engineering work itself is structured. At both Anthropic and Cursor, an increasing portion of engineering effort focuses on building environments where agents can execute more efficiently, rather than writing code directly. This suggests that future engineering roles may emphasize agent optimization, infrastructure design, and environment configuration over traditional software development.

The trend also reflects a larger pattern of AI spending optimization. As software engineers' AI consumption costs have grown substantially, platform teams are aggressively optimizing spend-per-token, making cloud-based execution more economical than local processing.

Meanwhile, a separate concern has emerged regarding how AI safety is being tested. Meta contractors posed as teenagers to test ChatGPT, Gemini, and other chatbots on high-risk topics including suicide, self-harm, eating disorders, and drug use, using 18 dummy accounts and submitting 3,748 prompts designed to assess safety guardrails. According to the report, OpenAI, Google, and Character.AI were unaware their systems were being tested in this manner, though Meta stated that such exercises are standard industry practice for ensuring safe and age-appropriate experiences.