findings

We poisoned our agent's memory on purpose. It healed itself.

Compound learning lets agents get smarter every run — but it keeps bad lessons forever, too. Here's how we made workflows verify their own learning, and the five experiment iterations it took to get the gate right.