The Six Layers of GenAI Engineering
Prompt, context, harness, loop, graph, and runtime engineering solve different problems in agent systems. Here is what each layer owns.
Millrace blog
Notes from the runtime.
Historical essays and current technical notes on graph engineering, long-running agents, and operational authority.
Long-Running Agents Don't Just Have a Memory Problem. They Have an Authority Problem.
Better memory helps an agent reason. Authoritative execution state determines what happened, what is legal, and what the system must do next.
The Research Is Independently Reconstructing Millrace
No single paper reproduces Millrace, but recent research independently supports nearly every major decision in its architecture.
Runtime Engineering Comes After Loop Engineering
Loop engineering is real, but once loops become infrastructure, the missing layer is runtime engineering.
Naive loops aren't the solution to loop engineering. Governed loops are.
Naive agent loops are useful, but reliable long-running agent work needs deterministic runtime judgment.