Anyone running AI agents in production knows the quiet frustration: the meter keeps running even when the agent is doing nothing at all, just sitting there waiting on a model to respond or a tool to finish. DigitalOcean thinks that waste has gone unaddressed for too long, and it built a new product around fixing it.
The company opened public preview this week for DigitalOcean Managed Agents, a set of services meant to give every AI agent its own isolated runtime, controlled access to outside tools, and fast inference, all bundled together instead of stitched from separate vendors. The pitch is less about making agents smarter and more about making them affordable to run at scale.
Vinay Kumar, DigitalOcean’s chief product and technology officer, compared the shift to the early cloud computing era, when renting a virtual machine on EC2 became the default starting point for building anything online. He argued agents now play that same role, except they need far more than raw compute. They need somewhere secure to execute untrusted code, a way to persist context across long and unpredictable sessions, and infrastructure that can pause and resume in milliseconds rather than the minutes a virtual machine typically takes.
That pause and resume detail turns out to matter more than it sounds. According to DigitalOcean’s own testing, its runtimes resume from a paused state in 305 milliseconds, which the company says beats competing offerings by 46 percent. Billing follows the same logic. CPU charges apply only while an agent is actively computing, not while it idles waiting on a response, which DigitalOcean claims can cut total costs by up to 37 percent compared with a leading sandbox competitor.
Early adopters are already testing the idea in production. Qencode built a support triage agent on the platform that reviews incoming requests across Slack, email, and Intercom, then routes tickets automatically, saving the team what its CEO estimated at four to eight hours a week. OpenHands and Amplitude are building similar agent workflows on top of it as well.
Whether idle-time billing becomes the norm across the industry remains to be seen. For now, DigitalOcean is betting that the biggest hidden cost of running AI agents was never the thinking. It was the waiting.
