Mubit and Anil on proprietary agent weights
Your agents are out here doing real work and most of it just evaporates into API logs, which is the part that got me. notanilp built Mubit to fix that: the output from those runs gets folded back into proprietary weights you actually own, and it keeps compounding. Traces from Langfuse, Braintrust, Logfire all feed in, and then smaller models get continually trained on the stuff you keep doing over and over. If you and a competitor are both building on the same frontier APIs, your intelligence layer was never really yours to begin with, so this is a very clever way around that. Credit to Anil and the team
