Tutorial 1 of 7
Review an AI agent before launch in 10 minutes
Beginner 10 min
Open the customer-support template, let Agentlab run five AI tools on it, then read the release verdict, the six areas and the findings that block the launch.
Agentlab is a workbench for one AI agent or LLM feature. Everything you know about it — its MCP configuration, a transcript, sample outputs, a RAG answer, the expected traffic — becomes a check. Every check runs one usefulweb.site tool, and Agentlab turns the results into findings, release gates and a report.
1. Open the template
Click Launch Agentlab. On the welcome card choose Customer-support agent: launch review (or File → Templates…). The project describes the support assistant of an online shop:
| Check | Kind | Tool |
|---|---|---|
| Agent MCP configuration | Agent configuration | Agent Permission Graph |
| Refund conversation (order 48213) | Agent trace | Tool Call Trace Explorer |
| Ticket triage JSON | Structured output samples | Structured Output Validator |
| Refund policy answer | RAG answer with citations | RAG Citation Debugger |
| Support traffic on OpenRouter prices | Cost scenario | LLM Cost Explorer |
Opening a template runs all checks right away. The status bar counts along (Running 3 of 5 …) and ends with Run all finished: 5 checks ran. Not ready: at least one gate fails.
2. Read the verdict
The Dashboard opens by itself. The banner at the top is the release verdict: Not ready: at least one gate fails and below it Gates: 1 pass, 4 fail, 0 without data. The chip in the toolbar says the same (4 gates fail) on every tab.
Below the banner, the tiles count the current findings: 5 high, 6 medium, 4 low and 1 passed check. Click a tile to filter the findings list; click it again to show everything.
3. Look at the areas
The six area cards group the checks: Security, Reliability, Output quality, Grounding, Cost and models and Reach. The coloured top edge shows the worst finding of the area. In this template:
- Security has 4 medium findings: the allow rule
mcp__helpdesk__*pre-approves every helpdesk tool, the refunds server keeps an API key in plain text, runs an unpinnednpxpackage and auto-approvesissue_refund. - Reliability is red:
issue_refundfailed twice andsearch_kbwas called three times with identical arguments — a loop. - Output quality: only 4 of 6 triage outputs match the schema (66.7 %).
- Grounding: 1 of 4 claim sentences is supported. The answer promises refunds within 5 business days while the cited chunk says 10.
- Cost and models is green:
openai/gpt-4o-minicosts about 102 USD per month for 4,000 requests a day. - Reach has no checks in this project.
Click an area to filter the findings to it.
4. Jump into a check
Every finding has a button with the check's name. Click Refund conversation (order 48213): the Workbench opens with the transcript, the lane chart of all eight tool calls (red boxes are failures, outlined boxes repeat an earlier call) and the table per tool.
Fix the input, press Run again (Ctrl+Enter) and the dashboard updates. Agentlab keeps every run, so the History tab later shows which findings your change resolved.
5. What next
- Gates (key 3) shows which threshold each gate uses and which check decides it — see the tutorial on release gates.
- Report (key 6) is the printable summary for the launch meeting.
- Your project is autosaved in this browser. Use File → Save project file (Ctrl+S) for a copy you keep.
Agentlab never calls your model or your agent: it analyses what you paste, so the same input always gives the same result.
Ready to try it?