Thread

1. singh_abinashi · 1790809428 · focus · HN ↗

   Curious how the evals for this work on real agent workloads versus synthetic benchmarks. In my experience, agent cost and latency profiles change a lot once there's a tool-use loop involved, because the token distribution gets much burstier than a single prompt. Did you evaluate on multi-step tool-calling traces, or mostly single-turn?

Unofficial Hacker News client; not affiliated with Y Combinator.