AI Agents Are Thirsty for Power, and Builders Should Care

The AI industry is moving from one-shot chatbot answers toward systems that plan, call tools, retry, and delegate work to other agents. That shift changes more than the user experience. It changes the physical bill behind every workflow.

What happened

In a report for WIRED, Molly Taft describes how agentic workloads can generate hundreds of smaller model prompts from one user request. A coding task that runs for hours, for example, may repeatedly re-prompt itself while building pages, menus, datasets, or other components.

The scale can become dramatic. WIRED reports that OpenAI recently described a swarm of more than 10,000 agents sending 2.7 million messages while working on a mathematics problem. The exact energy cost is difficult to calculate because frontier labs disclose little about the power used by agent workloads.

Meta is also planning for persistent personal agents. Its Muse system is described as giving each user a dedicated computer in the cloud that can continue working while the user is offline. That model turns AI from an occasional query into a background service with a continuous infrastructure footprint.

Why builders should care

For product teams, the immediate lesson is operational. Agent loops, parallel helpers, long context windows, and repeated tool calls can multiply inference cost without appearing as extra user requests. A workflow that looks like one button in the interface may be dozens or hundreds of model calls underneath.

That makes budgets and limits part of the product architecture. Builders should track cost per completed task, cap retries, route simple steps to smaller models, and stop work when progress stalls. The same telemetry should expose latency, token growth, tool-call counts, and failure recovery, not just final answer quality.

This connects to the broader serving question covered in our analysis of agentic AI production stacks. Efficient agents will not come from a model choice alone. They need orchestration that knows when to parallelize, when to summarize, and when to stop.

The infrastructure race is therefore also a workflow-design race. As agents become more capable, the winning systems will be the ones that deliver useful autonomy with bounded compute, visible resource use, and predictable failure behavior.