newsfilter
← Interviews
AI Engineer13 min

Every Correction You Give Your Agent Is Data - Brian Douglas, Paper Compute

Brian Douglas

Conference Presentation, Talk

Paper Compute engineer presents a framework for capturing AI agent "pushback" events through the open-source Tapes tool to preserve institutional knowledge and mitigate the risk of junior engineers failing to develop senior-level context. By fine-tuning a 4B parameter model on these labeled traces, the speaker demonstrates how to create a proprietary "Twin Reviewer" that simulates senior engineering feedback, effectively closing the review gap without relying on inaccessible large-scale direct preference optimization. This approach addresses the "vibe coding" trend by prioritizing structured learning and data isolation over public model training to prevent agents from reinventing existing organizational systems.