An agent investigating a production issue spends its time in bursts. It queries logs, runs a reproduction script, tests a fix, then waits on the model or on a human approving the pull request. During the waits it uses almost no CPU, but its working state has to stay alive so it can pick up where it left off.
Provisioned infrastructure bills the wait anyway. You rented an hour, you get an hour. DigitalOcean's Managed Agents launch this week meters CPU on actual consumption, so the charge falls to zero while the agent waits on tokens. Their worked example: a two-vCPU session averaging 25 percent utilization and peaking at 4 GB of memory for an hour costs $0.060 in CPU and memory charges, against $0.126 for the same allocated capacity billed provisioned.
The resume number is what makes pausing rational. Snapshot the session, stop the CPU and memory charges, resume to readiness in 305 milliseconds, and the next answer arrives in 2.43 seconds, comparable to the 2.47 seconds of a session that never paused.
I wrote the full walk through on the FinOps handbook: the rate card, the benchmark table where they disclose the one lane they lose, and the credential-brokering half of the cost story. The agent billing problem ↗