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

Description and use

Reducing AI agent cost with OpenLIT: from $0.0052 to $0.0002 per run, a 26× savings by eliminating redundant tool calls and switching to a cheaper model.

Problem

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

Outcome

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.

Workflow

• 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

Evidence

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

Source

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

Source