Choose the test for the job
| Your question | Start here | What to inspect |
|---|---|---|
| Did it answer correctly? | Scenario evaluations | Expected text, schema, semantic or judge-based criteria. |
| Can someone bypass the boundary? | Red teaming | Responses to scoped attack mutations and the observed failure. |
| What changes under load? | Stress testing | Latency, throughput, errors and declared concurrency. |
| Did the agent complete the task? | Agent workflows | Tool actions, policy decisions and independently verified state. |
| What does the saved evidence establish? | Workflow reports | Eligible outcomes, missing measurements and evidence-linked findings. |
CLI or SDK?
Start with the CLI for terminal and CI workflows. Use the SDK when evaluation belongs inside your application or test suite. Guardian Mode and framework adapters have their own setup and limitations; they are not a substitute for a security review.
Run with your stack
Configure OpenAI, Anthropic, Azure OpenAI, Ant Ling or the Vercel AI SDK adapter. Provider support and tool capabilities vary by integration. Keep credentials in your controlled environment.
Review provider setup and reproducibility guidance.
What does a pass mean?
It means a declared check passed for the inputs and configuration used. A passing text check does not establish task completion. A completed conversation does not establish a verified workflow outcome. Inspect the relevant criteria, execution record and limitations.