EvalGlass

Use case · Drift investigation

Behavior feels different. Where did it drift?

Drift is the change no code review catches — same code, different outputs. Ask your coding agent to read traces you have already exported, turn them into checks, and compare against your expected behavior. EvalGlass makes drift a delta you can read.

Inspect recent traces and tell me where behavior drifted from our expected checks.
You Inspect recent traces and tell me where behavior drifted from our expected checks.
Your coding agent · Claude Code plugin Looked at the exported traces you provided and built checks around them. Tone consistency drifted −0.09 and one tool-call pattern changed; grounding held steady. This is a comparison against your expected checks on the traces you exported — not a live feed. Verdict informational: it shows where to look, it does not gate anything.

What it checks

trace-derived cases vs. your expected behavior · tone · grounding · tool-call patterns. It uses traces you have already exported (OpenTelemetry / OpenInference / local JSON) — no SDK on the evaluation path (a live pull is a separate, experimental opt-in lane). It is a comparison you run, not continuous monitoring — and evalglass watch next lets you schedule it: one run against a baseline, a typed drift.json, then exit (built for cron or CI, not a resident daemon). It flags a regression only when the runs are comparable and the paired interval clears zero — otherwise the delta is within_noise; not_comparable or a missing baseline is reported as exactly that. The watcher adds no exit class and never promotes the baseline.

The scorecard

drift check · vs. expected behavior informational
Tone consistency0.81 −0.09
Groundingstable
Tool-call patternchanged

Verdict informational — a comparison on exported traces, not a live monitor. Illustrative example, not a measured result.

What it will not claim

EvalGlass does not watch your traffic. It reads traces you export and compares them to your checks — a deliberate run, not an always-on monitor. evalglass watch next can schedule that comparison as a one-shot on cron or CI, but it never watches live traffic — and a no-comparable-regression result never means quality is fine. No false green →

Ask your coding agent.

Evaluate my agentic app using EvalGlass.

Related

All use cases
every change moment
Scorecards
what each run hands back
Get the plugin
two commands in Claude Code
Docs
the detailed reference