Modern agentic systems increasingly rely on skills: installable packages of natural language and code that teach an LLM agent to perform a domain task. As skill repositories grow, developers need automated quality signals on every change, yet evaluation today is largely anecdotal: a developer asks an agent to "try the skill," watches a demo, and forms a subjective impression. This yields neither reproducibility across runs nor comparability across versions, and scales poorly to marketplaces where one regression can silently break dozens of downstream workflows. We present AEVAL (Agentic Evaluation), a CI-integrated framework that replaces this practice with a deterministic, reproducible test pipeline for agentic skills. Every skill change triggers a test event: the skill runs against a developer declared evaluation contract inside an automated executor, emitting a structured, evidence grounded quality signal that downstream CI can route on. A key ingredient is a structural separation between executor and grader, preventing a subtle but pervasive failure mode: an agent that silently self-corrects during execution and then grades its own patched outputs as passing. Our contributions are: (i) a deterministic, change-triggered evaluation protocol with per-skill contracts and per-run artifact schemas; (ii) a formalization of self correction bias as a distinct failure mode of naive agentic evaluators; (iii) an executor/grader separation with a first-attempt grading rule and explicit self-correction tracking; and (iv) a tiered, grounded evidence fix suggestion scheme (LV1 causal, LV2 quality) posted as inline merge-request comments. Validated on real skills in a production agentic stack across multiple agent SDKs, AEVAL converts spurious 100% pass rates into reproducible first-attempt fail signals with an auditable record of every executor fix.
Sawyer Zhang, Alexander Wang, Sophie Leics.CL cs.AI
End-to-end task-success is the dominant way to evaluate LLM agents, but one aggregate number tells you that an agent regressed, not where. We present layer-isolated evaluation: a deployed ordering agent is decomposed into a fixed taxonomy of layers (ontology, intent, routing, decomposition, escalation, safety, memory, and cross-cutting envelope/defense), each exercised by its own assertion slice in a deterministic, no-LLM "pure" mode. The pure suite (238 cases across 23 slices; 225 run in 2.39 s, ~10 ms/case) runs in CI on every change against a locked per-slice baseline. We validate by controlled regression injection, degrading one layer at a time across seven non-safety layers. The effect we did not design in is masking: the aggregate pass-rate barely moves (-1.7 to -5.9 pp for six local regressions), while the matching slice craters (-25 to -91 pp). A layer's slice reacting to its own fault is partly by construction; the measured results are (i) the aggregate masking and (ii) that damage stays off the other slices: the injected layer's slice is the single worst-hit in 5 of 7 cases and top-3 in 7 of 7 (mean rank 1.29 of 19). Localization replicates on a second, structurally different tenant (Starbucks SG): all seven matching slices crater, so it is not a single-catalog artifact. We position it as a concrete, deterministic instantiation of the component-level evaluation EDDOps prescribes but leaves unimplemented, with CheckList as ancestor and as the deterministic mirror image of whole-workflow stochastic mutation testing. Our contributions: (a) a fully decomposed, sub-second, no-LLM per-layer harness for a production agent, (b) a coverage-honesty test-adequacy criterion that refuses to score an unexercised layer, and (c) the regression-injection demonstration that per-slice baseline-locked gates localize regressions an aggregate metric masks.
Production teams deploying LLM chat agents face a specific quality assurance gap: existing evaluation tools test individual responses or simulate social interactions, but none systematically verify whether real users can achieve their goals through multi-turn conversation. We introduce a three-layer dogfooding framework that bridges this gap by combining canonical question-bank testing (Layer 1), random-walk multi-turn evaluation (Layer 2), and a goal-directed NPC (Non-Player Character) simulator with five structured goal types and a ten-category failure taxonomy (Layer 3). In a longitudinal case study on a production multi-agent system over roughly three months (257 evaluation runs; a 108-scenario NPC suite), we find that the three layers produce complementary regression signals: cross-layer correlation for response quality is weak within a synchronized run (Spearman rho between -0.15 and 0.14) and negative across the longitudinal series (rho down to -0.46), confirming that canonical correctness does not predict goal-directed conversation success. The NPC simulator achieves 77 percent goal achievement at 0.17 dollars per run (6,272x cheaper than human evaluation), enabling daily CI/CD integration with automated PROMOTE/HOLD/ROLLBACK release decisions. We release full prompt templates, the failure taxonomy, and a Python-first replicability guide so that other teams can adopt the framework for their own LLM chat agents.