What they found
The authors compared three evaluation frameworks on two analytics agents, using frozen execution traces (saved records of what the agents did). The three were trace-based LLM judging, text-only LLM judging with red-teaming, and deterministic heuristics (fixed rules). In Phase 1, 97% of configurations met the stability threshold, and stability patterns did not depend on the agent. In Phase 2, on 72 ground-truth test cases, the trace-based framework had the best balance of precision and recall (F1=0.90). Text-only evaluation matched it on recall but produced 5x more false alarms. In Phase 3, the study found 24 novel failures, with zero overlap between functional testing and security red-teaming.
Why it matters
Companies that put SI agents into multi-tool workflows need ways to catch failures before users do. This study suggests that what the evaluator can see matters a lot: access to the full trace beat text alone. It also suggests that functional checks and security checks find different problems, so one is not a substitute for the other. The authors also report that LLM-based evaluators found subtle errors that human reviewers missed. The abstract describes a controlled research setting with two agents, so how far this carries to other settings is not stated.
How it changes the race
This is practical, incremental guidance for teams deploying SI agents. It points to richer trace access and a split between functional and security testing as a way to cut missed failures and false alarms.
Safety and control
The finding that functional testing and security red-teaming found entirely different failures suggests that relying on only one leaves gaps. The authors recommend using both: trace-based evaluation for function and adversarial red-teaming for security.
The abstract, by the authors
Deploying LLM-based analytics agents in enterprise settings requires evaluation frameworks that can reliably detect failures across complex, multi-tool workflows. We present a three-phase comparative study of three evaluation frameworks, each representing a distinct evaluation paradigm (trace-based LLM judging, text-only LLM judging with red-teaming, and deterministic heuristics), applied to two analytics agents in a controlled research setting using frozen execution traces. Phase 1 quantifies evaluation harness stability through repeated measurements, finding that 97% of configurations meet the stability threshold and that stability patterns are agent-independent. Phase 2 measures known risk detection on 72 ground-truth test cases, where the trace-based framework achieves the best precision-recall balance (F1=0.90) while text-only evaluation matches on recall but generates 5x more false alarms. Phase 3 evaluates unknown risk discovery, uncovering 24 novel failures with zero overlap between functional testing and security red-teaming. Our results demonstrate that (1) the information available to the evaluation harness determines evaluation quality, (2) functional and security evaluation are complementary, and (3) LLM-based evaluation harnesses surface subtle errors that human reviewers miss. We recommend a two-framework strategy: trace-based evaluation for functional assessment and adversarial red-teaming for security validation.
Quoted from arXiv:2611.14493. The research and its words belong to its authors.
