EvalGlass

Use case · Agentic app evaluation

Does your tool-using agent do the right thing?

Evaluating an agentic app means checking behavior across steps — tool selection, trajectories, and outcomes. EvalGlass turns traces you have already exported into project-specific checks your coding agent runs locally, and returns a bounded scorecard you can gate in CI when you choose.

Evaluate my agent's tool use and trajectories against my expected checks.
You Evaluate my agent’s tool use and trajectories against my expected checks.
Your coding agent · Claude Code plugin Graded your support-agent runs at the trajectory unit, not one call at a time. Tool selection held; the task-outcome check dropped on the multi-step returns path, and one run came back non_evaluable because its trajectory was empty — never scored as 0.0. Verdict informational: an aggregate green here still would not prove the agent succeeded.

What it checks

Tool selection · step trajectories · task outcome · grounding and policy adherence — scoped to the unit you declare: a traces: route can set unit: trajectory|session|step next to grade the whole agent run, not one call. Checks are authored around your AI application’s real behavior, not a packaged agent benchmark.

Grade the whole run next

A single call is not the agent. A traces: route can declare unit: trajectory|session|step to grade the whole agent run, wiring the trajectory_shape@1 aggregate. Egress is worst-of-members (fail-closed) — the run is only as trustworthy as its weakest graded member. It is informational and cannot gate, and an empty trajectory is non_evaluable, never 0.0. An aggregate green is still not proof the agent succeeded.

The scorecard

trajectory run · support agent informational
Trajectory shape · unit: trajectory0.78 worst-of-members
Tool selection0.90 stable
Task outcome0.71 −0.08
Empty trajectorynon_evaluable

Verdict informational — graded at the trajectory unit; the empty run is non_evaluable, never 0.0, and nothing gates until you approve a threshold. Illustrative example, not a measured result.

Measured, not certified

EvalGlass measures what you defined and reports what it can claim. It never certifies the agent is safe or correct; authority to gate comes from host-owned records, and a fresh setup gates nothing.

What it will not claim

An aggregate green is not proof the agent succeeded. A whole-run grade rolls up the members it could measure — and a right answer can still come from the wrong tool or a skipped step. EvalGlass grades the observable trace evidence, never the model’s self-narrated reasoning, and nothing gates until you approve a threshold in your own config. No false green →

Ask your coding agent.

Evaluate my agentic app using EvalGlass.

In the docs

How it works →
the check loop, end to end
Tool-use & trajectory evaluation
how the checks read tool calls, steps, and recovery
Traces →
exported-trace import
CI for AI →
gate only when you approve
Scorecards →
what every verdict means