Feizi Claims Harness Optimization for Open-Weight Agents

Soheil Feizi stated that harness optimization enabled open-weight agents to reach frontier-level performance at roughly 3 percent of the inference cost. He described three models, each showing a distinct failure mode, and three separate fixes that addressed them. Feizi added that the harness proved more important than expected. The post comes from the account of the University of Maryland associate professor and RELAI founder.