Skills
Validating the plan against a fixed schema before any tool call is what makes an agent reproducible, not the retry and validation layers
arXiv 2608.26197 (2026-08-26) stacked finite-state control, forced tool selection, output validation, and bounded retries on two open-weight models and got mixed results across all four model-task cells. Adding structured planning, where the plan is checked against a fixed schema before any tool is invoked, pushed three of four cells to a Reproducibility Rate and Determinism Index of 1.000 at N=100 and task success to 100%. Token cost fell in every cell, but latency split by model, one sped up and one slowed markedly, so the trade has to be measured per model rather than assumed.
↳ Follow the thread