Evals & quality
Test your agent on real cases before you publish, so you ship the capabilities that actually work.
An eval is a test case for your agent. You give it an input and check the output. Evals turn "it felt right" into pass or fail you can see, so you know what your agent can actually do before anyone else uses it.
Why run evals
- Prove the capability. Confirm the agent handles the real cases it's meant to, not just the demo.
- Catch regressions. After a change, re-run your cases to make sure nothing that worked before is now broken.
- Find what to fix. A failed case shows you where the agent fell short, so you know which skill or tool to improve.
Cases and suites
- A case is one example: an input, and what a good answer looks like.
- A suite is a set of cases you run together, so you can check a whole capability at once.
The builder can draft cases for you from what the agent is meant to do, and you can add your own real examples.
Running evals
Run a case or a whole suite and watch the results stream in. Each case comes back as pass or fail, with the agent's output so you can see why. When a case fails, fix the skill or tool it depends on and run it again.
Quality and the publish gate
Kavela scores your skills against the base model. A skill earns its place only if it adds something the model couldn't already do on its own. This keeps the marketplace free of thin prompt-wrappers.
Before a paid agent can go live, it passes a readiness check. The check flags anything missing, such as an untested tool or a skill that doesn't add value, so you fix it before users see it.