We used OpenLIT to make an AI agent cheaper and safer, then shipped the walkthrough

Описание и применение

Сокращение стоимости работы AI‑агента с помощью OpenLIT: от $0,0052 до $0,0002 за запуск, 26× экономия за счёт устранения лишних вызовов инструментов и перехода на более дешевую модель.

Задача

High-cost, redundant tool calls in an AI support-desk agent leading to inefficiency and safety concerns.

Результат

Reduced per-run cost from $0.0052 to $0.0002 (26× cheaper) by eliminating redundant tool calls and switching to a cheaper model, while improving safety through better loop logic.

Рабочий процесс

• Scan the repo for issues
• Fix high-severity problems (missing docstrings, insecure paths)
• Run the agent and capture the trace with OpenLIT
• Analyze the trace to identify redundant tool calls
• Tighten the loop logic
• Switch to a cheaper model (gpt-4o-mini)
• Re-run and verify cost reduction
• Ship the improved agent

Доказательства

First run: repeat lookup_ticket / create_ticket calls, spend $0.0052.

Источник

Second run, after tightening the loop and switching to gpt-4o-mini: $0.0002, one lookup, one create.

Источник