Outcome-Only Evaluation Misses Half of Silent Agent Failures, New Research Shows

Current systems for evaluating language model agents rely almost entirely on checking whether the final answer is correct—a blunt approach that systematically misses failures hidden beneath apparently successful outcomes.

A new analysis from arXiv quantifies this blindness through controlled experiments on tool-using agents in a customer support scenario. Researchers injected deliberate faults at specific steps and measured whether different evaluation methods could detect them, distinguishing between failures that produced wrong customer-visible results ("loud" faults) and those that succeeded anyway despite internal errors ("silent" faults).

The Outcome-Only Problem#

Outcome-only judges—the current production standard—correctly identified 84% of loud faults but caught only 45% of silent ones, while incorrectly flagging 33% of actually correct trajectories as failures. This gap exists because these judges never look at how the agent reached its answer, only whether the final response was right.

The researchers tested their method on 400 trajectories using five different evaluation approaches: programmatic rules, outcome-only LLM judges, step-by-step rubrics at two different model scales, and an ensemble using self-consistency voting.

A Better Alternative: Step-Level Assessment#

A step-rubric judge that examined the agent's reasoning trajectory performed substantially better, reaching 77% recall on silent faults while maintaining zero false alarms—but at triple the computational cost. This method reads each step of the agent's work, not just the final reply.

Another surprising finding: every judge tested ignored the final response text entirely. Researchers appended a fabricated promise to an otherwise correct trajectory and found that outcome-only rules rejected it completely, while the step judge missed it 82% of the time. Self-consistency voting—repeating the evaluation multiple times and checking agreement—tripled costs without improving accuracy on any metric.

Implications for Agent Evaluation#

The researchers argue that evaluation frameworks must now stratify recall measurements by outcome survival, separating how well judges detect failures that users would notice from those buried within successful interactions. They have released their test environment, fault injection system, all raw verdicts, and analysis code to enable reproducible offline verification of their findings.