Today we are publishing our first collaborative research with
@Eon_io_
We gave an Eragon agent a readable guide to a company's systems, then let GEPA rewrite that guide from the agent’s own failed traces.
GPT-5.6 Luna is the cheapest model we tested, and cold it was the weakest of the GPT and Claude models: 7/35.
Post GEPA: 29/35.