Agent Quality Evaluation¶
This chapter covers the screen that measures response quality and safety in batches for built agents. The Agent Creation → Agent Quality Evaluation menu in the left sidebar is in scope.
For the single-input debugging flow inside the canvas, see Agent Operations · Execution and Debugging. This chapter focuses on batch evaluation — running many inputs at once and scoring them.
Accessing the Screen¶
Select Agent Creation → Agent Quality Evaluation in the left sidebar. The screen has two tabs at the top.
| Tab | Covers |
|---|---|
| 품질 평가 (Quality Evaluation) | Test execution history + start a new test (includes Harmbench safety evaluation) |
| 품질 척도 정의 (Quality Criteria) | Manage criteria presets used to score responses |
Common top-right actions: + New Test, Harmbench, Refresh.
Quality Evaluation Tab¶
Narrow the history with the top status filters (All / Running / Completed / Error) and the Search by agentflow box. When no tests exist, the body shows an empty state — "테스트 기록이 없습니다 — Select an agentflow and start a test" — with a Start New Test button.
Running a New Test¶
Clicking + New Test at the top right opens the Batch Test Creation modal. Upload a dataset file (questions + expected answers) to score many cases at once.
| Item | Description |
|---|---|
| Dataset format | .xlsx, .xls, .csv, .json (Harmbench / OpenAI and other common formats) |
| Column structure | At minimum question / expected_answer. Extra columns (category, tags) are used in result analysis |
| Upload method | Drag-and-drop or click-to-upload |
| Quick start | Use the Example or Excel buttons at the top of the modal to download a sample dataset |
After upload, choose the target agent and a criteria preset to run. Results return to the same history list and transition from Running → Completed with scores filled in.
Harmbench Safety Evaluation¶
The Harmbench button at the top right triggers an immediate safety evaluation using the public Harmbench dataset (Standard, 250 items) — no need to prepare your own dataset. It checks how the model reacts to harmful prompts against a standardized baseline.
| Item | Description |
|---|---|
| Source | HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusal |
| Dataset size | 250 items (Standard split) |
| Columns | behavior (harmful scenario prompt) / category (taxonomy) / behavior_id (identifier) |
| Result | Refusal rate, bypass rate, sensitive-response rate, etc. show up as safety scores on the history card |
Dataset contents
The Harmbench dataset contains sensitive prompts including chemical / biological / cyber risk scenarios. They are shown verbatim on screen — do not copy or redistribute beyond evaluation use.
Quality Criteria Tab¶
Manage criteria presets used to score responses. The solution ships official presets (e.g., Safety, Harmfulness, Ethics/Fairness, Expression Accuracy) and your own custom presets side by side as cards.
Creating a New Preset¶
Clicking + New Preset at the top right opens the preset authoring modal.
| Item | Description |
|---|---|
| Preset name | A one-liner identifiable to others (e.g., "Internal FAQ response quality") |
| Description | A paragraph about when this criteria applies |
| Category | Area classification — Safety / Accuracy / Tone / etc. |
| Criteria | Detailed metrics and weights used in scoring. Multiple can be registered per area |
Saved presets show up in the Criteria Preset dropdown of the Quality Evaluation → New Test modal.
Recommended Flow¶
- Define criteria first — Create the criteria preset that fits your organization before running tests, so all subsequent tests are comparable on the same scale.
- Start with a small dataset — Begin with ≤50 sample items to learn the criteria and result interpretation, then scale up.
- Run Harmbench regularly — Model or prompt changes can cause safety regressions; run it before every deploy.
- Share results with governance — For agents with "Org-wide" impact scope, use the evaluation score in AI Governance approval review.
Related Chapters¶
- Creating an Agent — Build the agent to be evaluated
- Agent Operations · Execution and Debugging — Quick single-input run
- AI Governance — Risk Assessment & Review — Use evaluation scores in approval flow
Contact¶
For Agent Quality Evaluation questions, please contact the Xgen Solution Administrator.