A few weeks ago, one of our developers pushed a refactor touching one of the agents Latitude runs in production. It looked harmless in review. Every request kept going through and latency stayed flat, but the prompt cache hit rate (how often an AI successfully reuses its cache instead of rereading information), which is normally above 80 percent, fell to 30. For perspective, cache reads cost about a tenth of what fresh input tokens do, so the drop tripled our effective token spend on the same traffic.
Thankfully, a monitor in Latitude was watching that rate and the drop opened an incident which, in turn, dispatched Claude Code in the cloud, and a few minutes later a pull request appeared that restored the cache structure the refactor had broken. A developer reviewed the change, merged it, and the hit rate recovered within the hour.
Latitude
Hey Product Hunt 👋 I’m César, founder of Latitude.
Over the past year, we’ve spoken with many teams running AI agents in production. Across all those conversations, one question kept coming up:
Is your agent getting better over time?
Most teams could answer for a few regression tests or isolated metrics, but not for the whole agent. Current eval systems have a coverage problem.
That’s why we built Agent Score.
Connect your production traces and Latitude automatically evaluates your agent across:
• Outcome
• Reliability
• Cost
• Speed
• Safety
It brings everything together into one quality score that updates daily, so you can see whether your agent is improving, what’s lowering its score, and the production sessions behind every result.
From there, Latitude helps you investigate recurring failures, track them over time, and send the evidence to your coding agent to open a fix.
We’d love for you to try Agent Score and tell us what you think. Any feedback is more than welcome 🙏
Mastra
oss ftw!
Can this detect issues that traditional logs and traces usually miss ?
@heycesr and @paula_cavero Congratulations on the launch of #AgentScore 🌟✌️. Quick Q, Is there any coding involved in the association of the production agent or is it more passive integration?
EverTutor AI
one daily score is nice but i'd worry it hides stuff. like if cost goes down but outcome quality drops a bit, the total could still look fine
can we set our own weights for each category? for us safety matters way more than speed
congrats on the 9th launch btw, that's wild