Tool discipline
Score whether each tool call reduced uncertainty and stayed within the task budget.
Agent evaluation platform
Agent failures usually appear before the final answer. Vectory evaluates whether the agent inspected the right state, used tools for a reason, recovered after errors, stayed in scope, and verified completion.
An agent evaluation platform should inspect every step of the run, not only the final message. That makes repeated loops, unsupported completion claims, weak recovery, and unsafe tool use visible before release.
Score whether each tool call reduced uncertainty and stayed within the task budget.
Check whether failures caused a changed action, a narrower hypothesis, and a later verification step.
Turn production misses into repeatable tests after redaction so future releases do not repeat them.