Switch the layers off one by one and you can see why agents collapse
The 12-layer stack, said to be distilled from experience with more than 400 production agents, is easiest to grasp by toggling each layer on and off as a boolean flag in a hypothetical emergency-room triage agent and reproducing the failures. Without the data-understanding layer, decisions get made with the allergy field missing; without context engineering, a 29-year-old and a 68-year-old patient get the same recommendation; and without the semantic layer, abbreviations like MI go unresolved. Loop engineering deals not with whether the model received the allergy information but with whether its output actually respected it, through a check → fail → correct → regenerate flow, while evaluation engineering looks not at fixing a single case but at the pass rate over a set of cases (22% → 100% in the demo). Harness engineering reproduces and traces the whole run, down to settings, model version, retrieved context and tool calls, and infrastructure engineering is about surviving external failures such as tool timeouts through retries, fallbacks and escalation. The point of the stack is that tools keep changing but the system rarely does.
Key points
- Context engineering asks whether the information was given; loop engineering, whether the output honored it
- Evaluation means case-level pass rates; the harness means reproducing and tracing the whole run
- The infrastructure layer covers external tools going down even when the agent's judgment is right
- Continual learning is a cycle: detect recurring failure patterns → propose fixes → human approval → regression tests
This is a note: my own write-up after reading several sources. It has no source link; events checked against a source are kept as records.