The switches in your agent workflow that nobody reads

Описание и применение

Многие переключатели в конфигурациях агентских рабочих процессов объявлены, проверены и задокументированы, но не читаются кодом, влияющим на поведение, что приводит к ложной уверенности и избыточной работе. Три‑шаговый аудит позволяет выявить и удалить или подключить такие «мертвые» переключатели, улучшая корректность и прозрачность процесса.

Задача

Many configuration switches in agent workflows are declared, validated, and documented but never actually read by any behavior‑changing code, leading to false confidence and wasted effort.

Результат

The audit revealed that most switches were only validated, not consumed. Only one knob actually changed behavior; others were either gated checks or inert. The process also uncovered a high proportion of duplicated reference cards in dispatched contexts, highlighting a separate cost issue.

Рабочий процесс

• List every config key from the schema or documentation.
• Grep the codebase for each key and classify hits as schema/enum, validator branch, metadata/pass‑through, or behavior branch.
• Adjudicate keys with zero behavior branches: wire them up or delete them.
• Document the decision and update the config accordingly.

Доказательства

A switch can be declared, documented, schema-validated, and still have zero effect, because validation only proves the declaration is well-formed, never that anything acts on it.

Источник

Step 2 — For each key, find its read points. Grep the codebase for the key name, then classify each hit: schema/enum definition, validator branch, metadata/pass‑through, behavior branch.

Источник

We ran this audit on our own protocol, which has six knobs that look like they tune how much process a task goes through.

Источник

46% of the bytes we shipped to our own agents were injected reference cards, and 38% of the total was the same card sent again.

Источник