Agent Internals — Page 4

Memory management, not reasoning capability, determines whether production agents actually scale.

Agents run code nobody vetted, so shared kernels are not enough.

Models forget everything, so the runtime must decide what to remember.

Borrowing OS scheduling primitives would have saved three of eleven agents.

Five agent loop designs, and which problems each one fits.

Agents fail silently, so one classic breaker can't protect them.