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

I built a small support-desk agent, scanned the GitHub repo, fixed the high-severity issues (missing docstrings, paths with no jail), ran it, and watched the trace.

First run vs second run

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

AI Analysis called out the waste.

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

That’s roughly a 26× cost drop on the same job, not from guessing, from reading the trace.

The loop OpenLIT is for

scan → trace → cost → tool misuse → ship a better agent

Without observability, the agent “worked.” With OpenLIT you see the waste: redundant tool calls, spend per run, and the exact place to tighten the loop before you ship.

If this is useful to you

• Use case : DM me or email [e-mail]. We’ll help you wire OpenLIT in.
• Feature : message us, or open an issue.
• Collaborate : connect with me on LinkedIn and we’ll figure something out.
• OpenLIT Enterprise beta (trial access) : same inbox. We’re taking design partners.

Links

• Video: https://youtu.be/mSOuJxW8cHs
• Repo (Apache 2.0): https://github.com/openlit/openlit
• Docs: https://docs.openlit.io

If the repo helps, a star actually does help others find it.