loading
Loading.Every claim we make about agents in production comes with the number we measured or the study someone else ran, and a way to check it. We ship the script before we ship the percentage. Run it on your own repo and get your own figure.
First-party, reproducible: ~60% fewer tokens per doc lookup, measured on 20 real queries against our own 145-doc repo. One command reproduces it on yours, and trovex measure reads your agent's real before/after.
read the evidenceFour independent 2025-2026 studies, read together: Berkeley's MAST failure taxonomy (1,600+ traces, 7 frameworks), ORAgentBench, an enterprise-scale study, and a concurrent-coordination one. The model is not the bottleneck.
read the evidenceGrounded in Berkeley MAST (arXiv:2503.13657) and ORAgentBench (arXiv:2606.19787, best agent 35.5% of expert tasks). Verification is one of the three failure categories the taxonomy names.
read the evidenceThe studies on whether agents finish complex work, summarized honestly with the numbers each one reports, not the headline anyone wished for.
read the evidenceIf you are rolling agents out across a team and the token bill or the disagreement is starting to hurt, we measure it on your stack and leave you a figure you can reproduce without us.