AI Engineer13 min
Every Correction You Give Your Agent Is Data - Brian Douglas, Paper Compute
Conference Presentation, Talk
- Future work aims to construct self-healing infrastructure and scale parallel agents to approximately 10 concurrent instances while offloading defined tasks such as documentation, test writing, and code cleanup to multiple agents in tandem.
- Teams are expected to shift toward reviewing code on the command line via AI agents rather than providing oversight directly on GitHub to mitigate code review pushback, utilizing session traces and labels to generate post-training datasets.
- A preferred approach involves creating "skills" to interact with trace data from tools like Tapes or Braintruss before attempting direct model training, while specialized fine-tuning on smaller GPUs like the 5090 is achievable by treating weights as "margin notes" for pre-trained models.
- Direct Preference Optimization (DPO) presents significant hurdles, requiring 70B+ parameter models and large hardware not typically available locally, often necessitating prepayment of substantial costs for services like Brev or Fireworks.
- Organizations can locally build harnesses to identify regressions and "golden sessions" without centralized platforms like Paper Console, with the Quinn 3.8B model suggested as a viable path for post-training scenarios requiring specific model sizes.
- Specific use cases include leveraging the Quinn model to evaluate Rust code against organizational coding standards, while the speaker anticipates that agents may currently lack context to recognize existing libraries, leading to unintended reinventions of open-source projects like mock libraries.
- Capturing session data enables the identification of "pushback" signals where human engineers possess specific knowledge, such as API structure, that agents lack, a trend that supports the evolution of "vibe coding" allowing software production without deep language understanding.