BenchFlow is a family of benchmark suites that put autonomous agents through real tasks — so you can measure skill, safety, and staying power before you ship.
Tests raw task competency across coding, reasoning, and tool use — the baseline for what an agent can actually do.
View suiteStress-tests agents that take real-world actions — probing for guardrail adherence when the stakes are non-trivial.
View suiteMeasures how well post-training methods hold up — tracking regressions and gains release over release.
View suite