Finding the AI Agents That Actually Matter with Leave-One-Out Ablation

Описание и применение

Библиотека agent‑ablation позволяет измерять влияние каждого агента в многоагентных системах, автоматически проводя leave‑one‑out аляцию и выявляя ключевых участников.

Задача

In multi‑agent AI pipelines, it is unclear which agents actually influence the final decision, leading to wasted compute and latency.

Результат

The library identifies which agents are load‑bearing, reports baseline verdict, per‑agent influence, load‑bearing ratio, and which removals change the outcome, enabling cost‑effective optimization of multi‑agent pipelines.

Рабочий процесс

• Collect findings from agents as Finding[]
• Compute baseline decision with decide(findings)
• For each finding: remove it, recompute decision, compare to baseline
• Aggregate results: identify load‑bearing agents, changes, and loadBearingRatio

Доказательства

The library reports: Baseline verdict, Per‑agent influence, Load‑bearing ratio, Which removals changed the final decision

Источник

In the benchmark, 6 of 9 automatically resolved cases collapsed to ‘escalate’ after removing a single specialist

Источник