Most agents almost work.
The hard part of an AI agent is not the first conversation. It is the second hundred — the ones where the agent does something subtly wrong and you do not notice for a week. This is built for the people who have shipped one and now want it to keep working.
You do not have to set anything up.
Sign in with GitHub. Pick the repository the agent lives in. Agent Etna reads the code: which model the agent calls, which tools it has, what its system prompt says it should do. There is no setup wizard because the work the wizard would do is already done.
You do not have to write the tests.
The system prompt is a description of the agent's job. Agent Etna treats it as a specification and generates the test conversations that exercise it — refusals, edge cases, the inputs the agent is supposed to handle and the ones it is supposed to push back on. The output is a count of what passed, what failed, and the actual conversation in each case. You do not need to know what an "eval" is to read it.
Warnings in sentences.
When the agent's behaviour drifts, the warning is not a stack trace. It is a sentence: "Your agent used to refuse risky requests. Now it tries to help. Here is the conversation that broke it." The proposed fix sits next to the warning, ready for you to read.
Nothing reaches your live agent without you.
Every fix lands as a real GitHub pull request you can read in a few seconds. Approve it and it merges. Close it and it does not. There is no autonomous-deploy toggle anywhere in the product, even when you are asleep. Removing the human from the loop removes the accountability with it.
Point it at your agent.
The free tier is live. Sign in with GitHub; the agent is auto-detected.