Can Agent Benchmarks Support Their Scores? Evidence-Supported Bounds for Interactive-Agent Evaluation

Researchers propose an outcome evidence reporting layer to improve the reliability of interactive agent benchmarks by explicitly tracking which runs have sufficient evidence of success versus uncertain cases. The framework evaluates five major AI benchmarks and reveals that surface-level outcome checks often fail to verify whether agents actually achieved intended results, making reported scores potentially misleading.

Agent benchmarking has become central to evaluating autonomous AI systems, yet this research exposes a fundamental flaw in how success is measured. Current benchmarks frequently rely on shallow outcome signals—checking whether an agent clicked 'Save' rather than verifying the actual state change occurred—creating a credibility gap between reported performance and genuine capability. This matters because inflated benchmark scores obscure real limitations in agent reliability, misleading developers and users about system readiness for production deployment.

The proposed evidence layer addresses this by introducing structured verification before scoring. Rather than producing a single aggregate number, the framework generates three-valued outcomes: Evidence Pass (verified success), Evidence Fail (confirmed failure), and Unknown (insufficient evidence). This approach transforms how uncertainty is handled, making invisible doubt explicit rather than hiding it in aggregate metrics. When applied to ANDROIDWORLD, AGENTDOJO, APPWORLD, tau3 bench retail, and MINIWOB, the framework revealed multiple distinct failure modes, suggesting that current benchmarks conflate different categories of agent failures.

For the AI development community, this has immediate implications for benchmark credibility and comparability. Teams using different benchmarks may be measuring fundamentally different things due to varying evidence standards. The framework's transparency about uncertainty aligns with broader calls for more rigorous AI evaluation practices. As enterprises increasingly deploy autonomous agents, the difference between claimed and evidenced capability becomes commercially significant. Future benchmark design will likely need to adopt similar evidence-based verification to maintain credibility as agent systems become more complex and mission-critical.

• →Current agent benchmarks often use surface-level outcome checks that don't verify actual state changes occurred, making scores unreliable.
• →A new evidence reporting layer assigns three labels (Pass, Fail, Unknown) rather than binary success/failure, making uncertainty explicit.
• →Testing across five major benchmarks revealed multiple distinct failure modes previously hidden in aggregate success rates.
• →The framework improves benchmark transparency without requiring modifications to existing tasks, agents, or evaluators.
• →Explicit evidence tracking could become standard practice as agent systems move into production environments where reliability matters most.