The numbers in that benchmark should stop every team running ReAct-style agents in their tracks: 90.8 percent of retries were spent on hallucinated tool calls. That is not a model quality problem. It is an architecture problem. Prompt tuning will not save you, and we agree. If your agent is repeatedly calling tools that do not exist or invoking arguments that cannot resolve, no instruction prompt will fix a structural flaw that is baked into the retry loop itself. The retry budget is being spent on the wrong failures, and that is a design decision, not an unfortunate accident.
What this means for you is straightforward: your agent is likely wasting time and compute on errors that were never going to succeed. The evidence points to three structural changes that eliminate wasted retries entirely. We are not going to relitigate the technical details here, but the practical takeaway is that you should stop treating retries as a safety net and start treating them as a diagnostic signal. If a retry is triggered, it should be because the model made a reasonable attempt that failed due to transient conditions, not because the tool-calling layer allowed an impossible request through in the first place. The architecture should be validating tool schemas, enforcing argument types, and rejecting hallucinated calls before they ever consume a retry.
The deeper lesson is about where you invest your optimization effort. Most teams pour energy into prompt engineering, tweaking system messages, and hoping the model "behaves better." But the data suggests that the model is not the primary failure point. The retry loop is. If you fix the loop, you reduce the pressure on the model to be perfect, which is a far more sustainable path to reliability. You are not asking the model to stop making mistakes; you are asking the system to stop paying for mistakes that are architecturally avoidable.
So, here is the concrete point: audit your retry logs today. Look at the actual errors that triggered retries in your last 200 tasks. If the majority are hallucinated tool calls, you have an architecture problem, not a prompt problem. Stop tuning prompts and start adding validation layers, schema checks, and fail-fast mechanisms. The retry budget is a finite resource, and every wasted retry is a lost opportunity to solve a real problem. The fix is not to retry harder; it is to make retries unnecessary.