Logistics · Agent debugging
Parcelly ships AI changes with evidence
How Parcelly found a retry loop tripling token spend on a single endpoint within a week of installing Cloudmind.
Results
- Reduction in monthly token spend
- 68%
- Time to find the root cause
- 1 week
- Retries per failing tool call
- 8 to 2
- Faster p95 agent completion
- 41%
The challenge
Parcelly runs a multi-step agent that reroutes shipments. Token costs had grown 3x in a quarter with no matching traffic growth, and nobody could explain it. The agent's own logs showed successful runs.
What they did
Nested tracing made the problem visible immediately: one tool call was failing schema validation, the agent silently retried it up to eight times, and the successful eighth attempt was all the logs recorded. A one-line schema fix and a retry cap cut spend back below the original baseline.
“The first week we installed it we found a retry loop that was quietly tripling our token bill on one endpoint. It paid for itself before the trial ended.”
Stop guessing whether your AI feature got better
Free for 50,000 traces a month. No credit card, no sales call, five minutes to your first trace.
Start free