Run the conversations your users will run.
Auto-generated test cases against your live agent. Real multi-turn flows. Adversarial prompts baked in. Pass/fail numbers you can ship to a stakeholder — not vibe checks.
What it does.
Test runs adversarial and happy-path conversations against your agent and measures how it behaves. Agent Etna builds a starter suite tailored to what your agent is for, then scores every reply on five things that matter — helpfulness, accuracy, safety, brand voice, conciseness. Failed cases get tracked so you can fix them, and the fix is verified against the same suite before it lands.
Auto-generated baselines.
Point Agent Etna at your agent's instructions and it produces a starter test suite tailored to what the agent is supposed to do. A coding-help agent gets coding tests. A customer-support agent gets escalation and refund tests. No template library to wade through.
Each test is a real multi-turn conversation, not a one-shot prompt — because that's how your users actually use the agent.
Adversarial coverage by default.
Every run includes adversarial probes — prompt injection, jailbreaks, data exfiltration, role confusion. So a green simulator doesn't just mean "the happy path works." It means the agent held up against the kinds of inputs production sees.
Aimed at the weak spots.
Agent Etna understands how your agent is put together and aims tests where it's most likely to break — the edges, the hand-offs, and the places real users find by accident.
Side-by-side comparisons.
Swap models (Claude vs. GPT vs. Gemini), swap system prompts, or swap parameter sets — Agent Etna runs the suite against both versions and shows you per-test deltas. "This change improved 12 tests, regressed 1, didn't move 47." A real number for a real meeting.
Run a baseline on your agent.
Free tier ships today. Connect a GitHub repo and your first test suite generates automatically.