An AI that suggests changes is easy to build. An AI whose changes you can trust needs evidence. Rylvo's improvement loops now prove a change before they offer it.
Agent Evolution finds the real cause
Agent Evolution reads every reply from Traces and shares one failure detector with Edge Cases, so both see the same problems. It traces a failure to its source, a prompt, a rule, a flow or knowledge, and proposes a fix there, saying exactly what it changed. A new rule is first checked against past replies before it can reach customers.
If you choose, it can do this by itself: draft the fix, apply it in Test and keep it only if the replies that failed now pass. Self-fixing stops at Test, and it is off until you turn it on. Rules that keep failing are rolled back.
Prompts that prove they are better
Self-improving prompts read the last two weeks of replies, suggest a few small edits and replay real conversations in Test, judged blind against the current version. An edit is offered only when it is clearly better and nothing gets worse. Replays can use the bot's read-only tools, and a fix found on schedule is announced to owners and admins.
You choose who judges
Pick the review model for Agent Evolution, Edge Cases, Tests and Red Team from a rated list. Red Team also lets you choose the model that writes attacks, tailored to the bot under test.
