Agentic AI and power budgets

Building AI agents that run 24/7 means treating energy as a first-class constraint.

We can talk about clever planning loops and tool-use all day, but if your inference stack eats watts like candy, your system won’t scale without carbon-intensive hosting. In this essay, I sketch a pragmatic budget:

Sensors -> Edge filtering -> Lightweight policy -> High-capacity model (sparingly)

Key tactics:

• Prefer small models with quantization on the hot path.
• Budget power like you budget memory: set hard caps per task.
• Push periodic heavy lifting to off-peak windows.

Callout: “Green by default” emerges from ruthless measurement. Instrument everything.

When the loop respects the budget, the system survives in the wild.