These five patterns cover common task shapes. Match your task to the shape, then adapt the
boundary, evaluator, promotion rule, and fallback to your domain.
Stacked semantic review
PR review agent
Best for
merge safety
Worker
LLM + tools
Evaluator
Deterministic tests and lint run first, then an LLM judge handles the semantic call the deterministic layer cannot make.
Target
output
Risk
medium / technically reversible, operationally costly
Together they cover generative and non-generative workers, reversible and irreversible
actions, selective autonomy, permanent gates, and evaluator maturation.