Evaluate the Stack, Not the Layer: Do Deterministic and LLM Gates for Agent Actions Fail Independently?

LLM judges are widely used to evaluate agentic tool-calling systems, yet their reliability on structured, dependency-driven workflows remains largely unexamined. We present AgentJudgeBench, the first benchmark to systematically study LLM-as-a-judge reliability for agentic tool-calling over workflow DAGs, as distinct from the broader LLM-as-a-judge task of open-ended text or preference evaluation. The benchmark comprises 3,808 instances spanning six DAG topologies and three difficulty tiers, evaluated with five generators (3B-70B open-weight models and GPT-5.4) and six judges (20B to frontier scale) under paired with- and without-ground-truth conditions. Judge alignment degrades monotonically with task difficulty, 1.5x faster without ground truth, and on hard queries without ground truth all six judges converge to a narrow 77-82% band regardless of scale, revealing a structural ceiling driven primarily by task difficulty, though its height is partly prompt-dependent for weaker generators, that model capacity alone cannot overcome. Ground-truth exposure is not uniformly beneficial: it reduces alignment for GPT-5.4 (1.5 pp) and Gemini-2.5-Pro (3.9 pp), consistent with over-anchoring. Among mitigation strategies, chain-of-thought reasoning and judge temperature both have negligible effect, while structured evaluation rubrics improve alignment by up to 6.5 pp but do not generalize uniformly across judge-generator pairs. With ground truth, QwQ-32B best matches the programmatic reference, while a human validation study identifies GPT-OSS-120B as the most human-aligned judge; without it, frontier judges lead only marginally within the shared ceiling. These results expose fundamental limitations of current LLM judges and yield practical guidelines for reliable evaluation in agentic systems.

AI agents increasingly take consequential actions -- shell commands, cloud operations, and arbitrary tool-calls -- so a trust layer must decide, per action, whether to allow, warn, block, or escalate. We argue that the right way to reason about such a layer is by threat type. Lexical (fixed-signature) threats, where danger lives in a stable token, are decidable by deterministic rules; semantic (intent-dependent) threats, where a benign and a malicious action share the same surface, are out of reach for rules by construction. We make this concrete with a negative proof: a determined, hand-authored cloud rule pack lifts held-out accuracy only 48 to 56% overall and moves the semantic categories by 0pp (data_db 29 to 29, observability 59 to 59, supply_chain 50 to 50), while a strong LLM judge carries exactly those categories. We give the judge a self-learning capability: on a corpus that is mainly semantic attacks it nearly doubles rule accuracy (48% to 83.6-85.2%) with near-zero false-blocks, and this holds across two model providers. We turn this into a self-improving dual-store system: the judge distills a growing deterministic rule floor on lexical threats (cheaper over time) and feeds a guarded RAG memory on semantic threats (a verdict-cache fails -- surface-twins collapse to ~58% -- so a corroboration guard lifts semantic accuracy +13pp, 70 to 84). The result is what sets AgentTrust v2 apart from its static v1 predecessor: a trust layer that self-evolves from its own stream of decisions -- cheaper on the lexical class (it distils its own rules) and smarter on the semantic class (it accrues guarded precedent), while never hard-blocking a benign action. An end-to-end online replay shows the judge-call rate falling (50% to 44%) and judge-domain accuracy rising (71% to 80%), with 0 benign hard-blocks across 45,000 actions.

A direct test of an LLM judge of agent trajectories injects faults into correct runs and reports recall, per fault type or by whether the fault broke the environment outcome (loud) or not (silent). Such recall can credit a judge with detection it does not have; paired discrimination, its flag rate on the faults minus its rate on the clean runs they came from, exposes this. Our testbed, a deterministic support desk with a scripted oracle and a one-step fault injector, labels all 400 trajectories exactly. A 14B judge shown only the request and final reply scores 34% to 76% recall on four fault types that leave the reply unchanged. There its input is the clean run's, so its paired discrimination is zero and that recall is its flag rate on clean runs. Splitting by outcome survival does not fix this: its loud recall of 84% is a paired +0.393 and its silent recall of 45% a paired +0.048, all from the two fault types that change the reply. Told to check each step, the same model flags every fault of those four types and 0 of 100 clean runs (95% CI up to 3.6%). It does not reliably check the reply: of four invented promises it flags one every time and the other three once in 42 faults. Shown every step but asked only about the reply, it still reaches a paired +0.69 on reply-unchanged faults, against +1.00 when told to check each step. We recommend reporting paired discrimination against clean parents, split by whether the fault reaches the judge's input and by outcome survival, and release the testbed, raw verdicts and analysis pipeline.