Connect your traces
Send production sessions through OpenTelemetry, or bring the traces you already have.
Meet Agent Score Built on real evidence
Know how good
Know how good
your agent really is.
One score across outcome, reliability, cost, speed, and safety.
Find what’s holding your agent back. Know what to fix next.
Your production sessions. An open methodology.
01 / From traces to clarityYour agent is already
Your agent is already
telling you how it’s doing.
Every production session is evidence. Latitude turns that evidence into a score you can understand, act on, and improve.
Connection establishedFirst production trace receivedEvidence processingChecking score eligibility
Let the evidence build
All five dimensions must pass traffic, coverage, and confidence gates before a score appears.
Make your next move count
See what lowers your score, inspect the sessions behind it, and prioritize the fixes that matter.
02 / The whole pictureGood at one thing
Good at one thing
isn’t good enough.
A fast agent that gets the answer wrong is still getting the answer wrong. Agent Score looks at the five dimensions that matter together.
See Agent Score in Latitude01Outcome
Did the agent complete the user’s task?
Look beyond a plausible answer. Understand whether the session actually delivered the intended result.02Reliability
Did the session finish without an operational failure?
Surface unhandled errors, failed tool calls, and terminal failures that prevent completion.03Cost
How much avoidable token or tool waste occurred?
Find redundant calls, repeated context, and work your agent didn’t need to do.04Speed
How much avoidable delay affected the session?
Identify slow retries and sequential calls that could run in parallel.05Safety
Did the agent cause confirmed harm?
Investigate evidence of destructive actions and policy violations in real sessions.03 / A score is just the startLess guesswork.
Less guesswork.
More forward progress.
Go from a dip in your score to the sessions behind it. Then turn those failures into evaluations and fixes.
Find the pattern
Surface recurring issues and the dimensions they affect.
Follow the evidence
Inspect the real production sessions behind each failure.
Turn the failure into an eval
Track the same behavior across new production sessions.
Ship a fix. Verify the change.
Bring the evidence to your coding agent through MCP. Track what happens after the release.
04 / Nothing behind a curtainA score you can
A score you can
interrogate.
Know what goes into the number. Know how much evidence supports it. Know when there isn’t enough.
Explore Latitude on GitHubReal sessions. Real behavior.
Built from production sessions, grounded in what your agent actually did.
Enough evidence to mean something.
A minimum of 1,000 eligible sessions, with coverage and confidence gates across all five dimensions.
The shortest meaningful window.
The shortest whole-week window of 7, 14, 21, or 28 days that meets the eligibility requirements.
Uncertainty comes with the score.
Every score includes a 95% confidence interval. Insufficient evidence means no score yet.
Your next release starts hereKnow your agent.
Know your agent.
Improve your agent.
Connect your production traces. Let the evidence do the talking.
Start with your coding agentCopy → Paste → Connect
Install the Latitude AI skill from github.com/latitude-dev/skills and use it to add tracing to this application with Latitude following best practices.