Preprint reports lower agent costs through automated task adaptation

Researchers report that AutoOnboarder, a system that adapts AI agents to specific tasks, matched or exceeded a Gemini 3.5 Flash baseline at up to 29× lower costs in tests involving genomic-sequencing prior authorization forms. A supervising coding agent adjusts the underlying model, prompt and software that mediates the agent's observations and actions, using tens of examples. Evaluation data stays on a separate virtual machine to prevent leakage. The preprint says the resulting agents remain imperfect and final performance varies greatly between onboarding runs.