Lean Agents

The problem

#

AI agent frameworks are built for prototyping, not production. A typical Python agent pod consumes 200–500 MiB. At that footprint, 100 agents need 50 GiB. Meanwhile, an ESP32 has 520 KiB total — Python can’t even fit.

The insight

#

AI coding assistants eliminated the productivity gap between compiled and interpreted languages. The tradeoff that justified heavy runtimes no longer exists — but the 20–50× resource penalty remains.

☁️

Scale Out

Kubernetes · Cloud
Memory cost × 1000s of pods

Scale Down

Microcontrollers · Edge
Physical RAM limits (KiB)