Testronaut Research
Experiments from Mission Control
Public methodology, normalized measurements, caveats, and results from experiments used to shape Testronaut.
Agent optimizationOctober 2026
Jev optimization experiments: guardrails and model routing
Two blocked experiments covering 70 aggregate runs, 210 mission-file attempts, and 120 isolated mission executions. The strongest result: guardrails can improve the trajectory without switching the execution model.
Explore research →