The emerging survey literature on agentic AI for UAVs makes it clear that “AI agents for drone image analytics” is no longer a single pattern but a family of architectures, each carving up perception, reasoning, and control in different ways. Sapkota et al. introduce the term “Agentic UAVs” to describe systems that integrate perception, cognition, control, and communication into layered, goal-driven agents that operate with contextual reasoning and memory, rather than fixed scripts or reactive control loops. arXiv.org arXiv.org alphaxiv.org In their framework, aerial image understanding is only one layer in a broader cognitive stack: perception agents extract structure from imagery and other sensors; cognitive agents plan and replan missions; control agents execute trajectories; and communication agents coordinate with humans and other UAVs. This layered view is useful when we start thinking about agentic frameworks as “judges” for benchmarking: the judging capability can itself be an agent, sitting in the cognition layer, consuming outputs from perception agents and workload metadata rather than raw pixels alone. arXiv.org alphaxiv.org 

Within this broader landscape, vision–language–driven agents are a distinct subclass. Sapkota et al. explicitly highlight vision–language models and multimodal sensing as key enabling technologies for Agentic UAVs, noting that they allow agents to parse complex scenes, follow natural-language instructions, and ground symbolic goals in visual context. arXiv.org alphaxiv.org These agents differ from traditional planners in that they can reason over image and text jointly, which makes them natural candidates for roles like “mission explainer,” “anomaly triager,” or, in our case, “benchmark judge” for aerial analytics workloads. Instead of judging purely from numeric metrics, a vision–language agent can look at a drone scene, read a workload description, inspect candidate outputs, and form a qualitative judgment about which pipeline better captures the intended analytic semantics. 

UAVCodeAgents by Sautenkov et al. provides a concrete multi-agent realization of this vision–language–centric approach for UAV mission planning. arXiv.org arXiv.org arXiv.org Their system uses a ReAct-style architecture where multiple agents, powered by large language and vision–language models, interpret satellite imagery and high-level natural language instructions, then collaboratively generate UAV trajectories. arXiv.org arXiv.org arXiv.org A core feature is a vision-grounded pixel-pointing mechanism that lets agents refer to precise locations on aerial maps, and a reactive thinking loop that enables iterative reflection, goal revision, and coordination as new observations arrive. arXiv.org arXiv.org In evaluation on large-scale fire detection missions, UAVCodeAgents reaches a 93% mission success rate with an average mission creation time of about 97 seconds when operated at a lower decoding temperature, illustrating that a team of reasoning-and-acting agents, anchored in visual context, can deliver robust, end-to-end behavior. arXiv.org arXiv.org arXiv.org While their agents are designed to plan rather than judge, the architecture is the same kind we would co-opt for an evaluative role: a vision–language agent that can “look,” “think,” and “act” by querying tools or recomputing metrics before rendering a verdict. 

Across these works, we can roughly distinguish three archetypes of agents relevant to drone image analytics. First are perception-centric agents, effectively wrappers around detection, segmentation, or classification models that expose their capabilities as callable tools within an agentic framework. arXiv.org alphaxiv.org Second are cognitive planning agents, like those in UAVCodeAgents, which translate goals and visual context into action sequences, refine them through ReAct loops, and manage uncertainty through deliberation. arXiv.org arXiv.org arXiv.org Third—more implicitly in the surveys—are oversight or monitoring agents that track mission state, constraints, and human guidance, and intervene or escalate when anomalies arise. arXiv.org arXiv.org For ezbenchmark, the “judge” fits best in this third category: an oversight agent that does not control drones directly, but evaluates analytic pipelines and their outputs against goals, constraints, and visual evidence, possibly calling perception tools or re-running queries to validate its own judgment before scoring. 

Agentic surveys also emphasize the role of multi-agent systems and collaboration, which is directly relevant to how we might structure an evaluative framework. arXiv.org alphaxiv.org Instead of a single monolithic judge, we can imagine a committee of agents: one agent specialized in geospatial consistency (checking object counts, extents, and spatial relations); another focused on temporal coherence across flights; another on narrative quality and interpretability of generated reports; and a final arbiter that aggregates their recommendations into a final ranking of pipelines. Sapkota et al. note that multi-agent coordination enables UAV swarms to share partial observations, negotiate tasks, and adapt to dynamic environments more effectively than single-agent systems. arXiv.org alphaxiv.org Translated into benchmarking, multi-agent evaluation would let different judges stress-test different aspects of a pipeline, with the ensemble acting as a richer, more discriminative “LLM-as-a-judge” than any single model pass. 

What makes this particularly attractive for an ezbenchmark-style adaptation of TPCH is that the agentic literature already leans heavily into reproducibility and benchmarking. UAVCodeAgents, for example, is explicitly released with plans for an open benchmark dataset for vision–language-based UAV planning, making their evaluation setup a template for standardized mission-level tasks and metrics in an agentic setting. arXiv.org arXiv.org Sapkota et al. argue for a “foundational framework” for Agentic UAVs that spans multiple domains—precision agriculture, construction, disaster response, inspection—and call out the need for system-level benchmarks that assess not only perception accuracy but also decision quality, mission flexibility, and human–AI interaction quality. arXiv.org arXiv.org This is very close in spirit to a TPCH-style workload benchmark, except operating at the level of missions and workflows rather than isolated queries. If we treat each ezbenchmark workload as a “mission” over a drone scenes catalog, an agentic judge can be evaluated on how consistently its preferences align with human experts when comparing alternative pipeline implementations for the same mission. 

In practice, using these agent types as judges means giving them access to more than just model outputs. An evaluative agent would see raw or tiled imagery, structured detections from classical or neural perception models, SQL outputs over our catalog, and the natural-language description of the analytic intent. It could then behave much like a planning agent, but in reverse: instead of generating a mission, it generates probes—additional queries, spot checks on specific tiles, sanity checks on object distributions—that help it decide which pipeline better fulfills the workload semantics. This is exactly the kind of “Reason + Act” loop that UAVCodeAgents demonstrates, only the action space is benchmark tooling instead of flight waypoints. arXiv.org arXiv.org The survey of Agentic UAVs suggests such introspective, tool-using behavior is central to robust autonomy in the field; using it in a judging capacity extends the same philosophy to benchmarking, pushing ezbenchmark beyond static metrics toward a living, agent-mediated evaluation process. arXiv.org arXiv.org alphaxiv.org 

Seen through this lens, enhancing ezbenchmark with an agentic judge is less about bolting on a new feature and more about aligning with where UAV autonomy research is already heading. Agentic UAV surveys formalize the components we need—perception tools, cognitive controllers, communication layers—and UAVCodeAgents shows how multi-agent ReAct with vision–language reasoning can reach high reliability on complex aerial tasks. arXiv.org arXiv.org arXiv.org arXiv.org Our benchmark can exploit those same design patterns: treat specialized detectors and SQL workloads as tools, wrap them in agents that can look, think, and act over drone imagery and metrics, and then measure how well those agents serve in an evaluative role. In doing so, ezbenchmark evolves from a TPCH adaptation into a testbed for agentic judgment itself, letting us benchmark not only pipelines, but also the very agents that will increasingly mediate how humans and UAVs reason about aerial imagery. 

Our References besides citations above: 

 Vision-LLMs-as-a-judge for aerial drone analytics benchmark

The idea of using vision‑LLMs and broader “vision‑language‑action” systems as judges in drone image analytics sits right at the intersection of two trends: treating multimodal models as evaluators rather than only solvers and pushing benchmarking away from narrow task metrics toward holistic, workload‑level quality. In the vision‑language world, this shift is now explicit. The MLLM‑as‑a‑Judge work builds a benchmark expressly to test how well multimodal LLMs can rate, compare, and rank outputs in visual tasks, not by their own task performance, but by how closely their judgments track human preferences across scoring, pair comparison, and batch ranking modes. mllm-judge.github.io That framing is exactly what we want for ezbenchmark: instead of only asking, “Did this pipeline’s SQL answer match the ground truth?” we also ask, “Given two pipelines’ outputs and an aerial scene, which better serves the analytic intent?” and let a vision‑LLM or VLA agent sit in that adjudicator role.

The details of MLLM‑as‑a‑Judge are instructive when we think about designing a drone analytics benchmark around an LLM judge. They construct a dataset by starting from image–instruction pairs across ten vision‑language datasets, collecting outputs from multiple MLLMs, and then building three evaluation modes: direct scoring of a single response, pairwise comparison between responses, and batch ranking over multiple candidates. Github Github Human annotations provide the reference signal for what “good judgment” looks like, and the final benchmark includes both high‑quality and deliberately hard subsets, with hallucination‑prone cases explicitly marked. Github mllm-judge.github.io Github When they run mainstream multimodal models through this setup, they find something specific: models align well with humans in pairwise comparisons but diverge much more in absolute scoring and in ranking whole batches of outputs. mllm-judge.github.io In a TPC‑H‑inspired drone benchmark, that suggests leaning heavily on pairwise “A vs B” judgments when using a vision‑LLM to compare query plans, detectors, or post‑processing pipelines on the same scene, and treating absolute scores more cautiously.

The same study surfaces a second lesson that matters a lot if we want to let model “grade” aerial analytics: multimodal judges are biased, hallucinate, and can be inconsistent even on the same input, including advanced systems like GPT‑4V. mllm-judge.github.io In the MLLM‑as‑a‑Judge dataset, they document systematic discrepancies between model and human preferences, instability of scores across reruns, and failure cases where the judge fixes superficial cues rather than substantive quality differences. mllm-judge.github.io Translated into the ezbenchmark world, we would not use a vision‑LLM judge as the sole source of truth for whether a drone pipeline is “correct.” Instead, we wrap it in the same discipline TPC‑H brought to SQL: ground metrics from the schema and query semantics, but let the judge operate in the “preference layer” to compare pipelines on interpretability, usefulness, anomaly salience, or robustness under slight perturbations of the workload. In other words, the judge augments but does not replace hard ground truth.

Where vision‑language‑action models become interesting is in how they extend this judging role beyond static scoring into interactive critique and tool use. A pure vision‑LLM judge can say, “Output B is better than output A for this aerial scene because it correctly flags all bridges and misses fewer small vehicles.” A VLA‑style judge can, in principle, go further: given the same scene and candidate outputs, it can call downstream tools to recompute coverage metrics, probe the pipeline with slightly modified prompts or thresholds, or even synthesize adversarial test cases and then update its assessment based on those active probes. Conceptually, we move from “LLM as passive grader” to “LLM‑agent as audit process” for drone analytics: an agent that not only scores, but also acts—running additional queries, zooming into tiles, checking object counts against a catalog—to justify and refine its judgment. The core evidence from MLLM‑as‑a‑Judge is that even in the static setting, models are more reliable in relative judgments than absolute ones; adding actions and tools is one way to further stabilize those relative preferences by grounding them in measurements instead of impressions. mllm-judge.github.io

For ezbenchmark, which already borrows TPC‑H’s idea of a fixed schema, canonical workloads, and comparable implementations, the natural evolution is to layer a multimodal judge on top of the existing quantitative metrics. Each workload instance can produce not only scalar metrics—latency, precision, recall, cost—but also rich artifacts: heatmaps, bounding box overlays, textual summaries, or “top‑k anomalies” for a given aerial corridor. We can then construct a secondary benchmark where a vision‑LLM or VLA agent receives the original drone imagery, the analytic intent (in natural language), and two or more candidate outputs, and must perform pairwise and batch comparisons analogous to MLLM‑as‑a‑Judge’s design. Github mllm-judge.github.io Human experts in drone analytics label which outputs they prefer and why, giving us a way to measure how often the judge agrees with those preferences, where it fails, and how sensitive it is to prompt phrasing or context. Over time, this gives ezbenchmark a second axis: “human‑aligned analytic quality” as seen through a vision‑LLM judge, sitting alongside traditional task metrics.

The last step is to close the loop. Once we have a calibrated vision‑LLM or VLA judge whose behavior on aerial scenes is profiled against human preferences, that judge can become part of the development cycle itself: ranking alternative detector ensembles, scoring layout choices in dashboards, or evaluating the narrative quality of auto‑generated inspection reports before they go to humans. The MLLM‑as‑a‑Judge results caution us to design this carefully—lean on pairwise comparisons, monitor bias, and keep human‑labeled “hard cases” in our benchmark so we can see where the judge struggles. mllm-judge.github.io But they also validate the basic premise that multimodal models can meaningfully act as evaluators, not just solvers, when anchored by a benchmark that looks a lot like what we are building: standardized visual tasks, structured outputs, and a clear notion of what “better” means for an analyst. In that sense, extending ezbenchmark from a TPC‑H‑style workload harness into a platform that leverages LLM‑as‑a‑judge for drone imagery is not a speculative leap; it is aligning our benchmark with where multimodal evaluation research is already going, and then grounding it in the specific semantics and stakes of aerial analytics.