Notes
Longer pieces on inference systems, agent architecture, and the infrastructure underneath the models.
- Anatomy of an inference requestWhat happens between hitting enter and seeing the first token, and why the two numbers that describe it are set by completely different machinery.
- The agent loop in productionThe loop itself is trivial. Everything that makes an agent good or bad in production lives in the parts the diagram hides.
- Context engineering is memory managementThe context window is not a text box. It is scarce, contended memory, and the old disciplines of managing scarce memory apply almost unchanged.