In the aerial drone analytics space, the comparison between plain vision‑LLMs, ReAct‑style agents, and broader agentic frameworks is about how we want our system to behave under pressure: do we want a single powerful model that “understands” scenes, or an ensemble of agents that can plan, probe, and correct themselves over time. The recent UAV‑CodeAgents work is a clean illustration of the second camp. It builds a multi‑agent framework on top of large language and vision‑language models, using the ReAct (Reason + Act) paradigm to interpret satellite imagery, ground high‑level natural language instructions, and collaboratively generate UAV trajectories. A vision‑grounded pixel‑pointing mechanism lets agents refer to precise locations on aerial maps, while a reactive thinking loop supports iterative reflection and dynamic goal revision in evolving environments. Evaluated on large‑scale fire detection missions, this ReAct+agentic stack achieves a 93% mission success rate with an average mission creation time of 96.96 seconds at a low decoding temperature, demonstrating that structured, multi‑step reasoning and tool use can deliver high reliability without blowing up latency.  

 By contrast, pure vision‑LLMs, even when multimodal and fairly large, tend to be evaluated on perceptual or question‑answer tasks rather than mission‑level outcomes. Sapkota and colleagues’ broader work on multimodal LLMs in domains like agriculture underscores the pattern: general‑purpose or domain‑adapted vision‑language models excel at flexible perception and instruction following, but their performance is typically reported as accuracy on classification, detection, or description benchmarks, not as end‑to‑end success in complex workflows. dblp In a benchmarking context like ezbenchmark, which is inspired by TPC‑H’s workload‑centric philosophy, that distinction matters. A vision‑LLM can certainly answer “What structures do we see in this tile?” or “Which parcels are likely orchards?” with impressive zero‑shot competence, but those answers are rarely tied directly to operational metrics like “Did the mission achieve its analytic goal without re‑flight?” or “How many follow‑up queries or human corrections were needed?” The agentic literature, especially around UAVs, starts from those operational questions and works backward to architecture.  CatalyzeX

 The Agentic UAVs survey by Sapkota, Roumeliotis, and Karkee makes that shift explicit. They define Agentic UAVs as systems that integrate perception, decision‑making, memory, and collaborative planning to operate adaptively in real environments, with goal‑driven behavior and contextual reasoning as first‑class design targets.  CatalyzeX In their taxonomy, vision‑LLMs and other multimodal models are enabling technologies inside a larger agentic stack rather than the entire solution. Perception components transform aerial imagery and other sensor data into structured representations; cognitive agents plan and replan missions; control agents execute actions; and communication agents manage interaction with humans and other UAVs across domains like precision agriculture, construction, disaster response, environmental monitoring, and inspection.  From an effectiveness standpoint, the survey argues that these agentic stacks surpass traditional UAV autonomy by improving mission flexibility, learning capacity, and system‑level robustness, but they also incur more architectural complexity. For a benchmark like ezbenchmark or a spatio‑temporal query engine like SpatialSky, this implies that evaluating “just the vision‑LLM” only tells part of the story; we also want metrics that capture how an agentic wrapper uses perception, memory, and planning to deliver reliable analytics over time.  CatalyzeX

 UAV‑CodeAgents sits at the intersection of these ideas and gives we quantitative hooks to work with. It exemplifies a multi‑agent ReAct framework where each agent is powered by an LLM or VLM but constrained by a structured action space: interpret imagery, reference map locations, synthesize mission code, revise plans, and coordinate with peers. The authors show that fine‑tuning Qwen2.5‑VL‑7B on 9,000 annotated satellite images substantially improves spatial grounding, which is a direct nod to the strength of vision‑LLMs as perception cores.  Yet the headline numbers—93% success rate, roughly 97‑second planning times—are achievements of the full agentic system, not the VLM alone.  If we imagine swapping that ReAct framework into an ezbenchmark workload, the effectiveness metrics we would record are not only pixel‑ or object‑level accuracies but also how many reasoning–action iterations the agents need to converge, how often they recover from ambiguous instructions without human help, and how consistently they satisfy constraints akin to TPC‑H’s query semantics when operating over a drone scenes catalog.

 The broader survey of Agentic LLMs reinforces why that ReAct pattern has become so central. It distinguishes between “plain” LLM use—where the model simply maps prompts to outputs—and agentic use, where LLMs plan, call tools, manage memory, and interact with other agents in pursuit of goals. alphaxiv.org UAV‑CodeAgents is explicitly cited as an example of this agentic turn in UAV mission planning: multi‑agent ReAct plus vision‑language grounding yields scalable, autonomous mission generation with minimal supervision. alphaxiv.org When we transfer that lens back to benchmarking, we get a natural three‑way comparison. Pure vision‑LLMs are cost‑effective for single‑step perception and natural language querying; ReAct frameworks wrap those models in explicit “think–act–observe–think” loops that can interrogate data and tools; full agentic UAV architectures, as surveyed by Sapkota et al., extend this further by embedding ReAct‑like cycles into a distributed system that includes collaboration, persistent memory, and multi‑mission learning.  alphaxiv.org  Each step up the ladder tends to increase implementation cost and complexity but also improves mission‑level robustness and adaptability in domains that look a lot like the use cases in SpatialSky and what we are sketching in ezbenchmark—multi‑tile analytics, evolving spatio‑temporal queries, and feedback‑driven missions over large areas.

 For the specific kinds of workloads in ezbenchmark and SpatialSky—workload chains over a spatial schema, spatio‑temporal pattern detection, and comparative evaluation of alternative pipelines—the existing literature suggests a division of labor rather than a straight winner. Vision‑LLMs, especially when domain‑tuned like the Qwen2.5‑VL‑7B variant in UAV‑CodeAgents, serve as powerful perception and explanation modules, mapping imagery and schema‑level metadata into natural language and structured hints.  dblp ReAct frameworks, exemplified by UAV‑CodeAgents, convert that perception into iterative planning and tool use, achieving high mission success and bounded planning time.  Agentic UAV architectures, as surveyed by Sapkota and colleagues, frame everything as part of a larger ecosystem where agents can accumulate experience, coordinate across missions, and adapt to new tasks and domains.  CatalyzeX If we encode those three regimes as configurations in ezbenchmark—vision‑LLM only, vision‑LLM+ReAct controller, and full agentic stack—we can attach metrics that reflect what the literature actually measures: task‑level accuracy and descriptive quality for the VLM, convergence behavior and mission‑success rates for ReAct, and cross‑mission adaptability and system‑level robustness for the agentic frameworks.  alphaxiv.org  CatalyzeX

 In that sense, incorporating ReAct and agentic metrics into ezbenchmark is less about chasing a trend and more about turning the UAV and agentic AI survey results into concrete benchmark dimensions. UAV‑CodeAgents gives us  model of how to quantify ReAct‑based mission planning performance in aerial scenarios, including success rates and planning time under different reasoning temperatures.  The Agentic UAVs survey gives us a taxonomy of capabilities—goal‑driven behavior, contextual reasoning, collaborative planning—that we can translate into workloads and evaluation criteria at the analytics level.  CatalyzeX And the broader Agentic LLMs perspective explains why simply swapping in a bigger or better vision‑LLM will not give us the same system‑level behavior as a ReAct or agentic framework; what matters is how the model is embedded in a loop of reasoning, action, and feedback. alphaxiv.org Together, they give us a roadmap for evolving ezbenchmark from a TPC‑H‑inspired catalog of queries into a testbed that can meaningfully compare vision‑LLMs, ReAct controllers, and full agentic UAV stacks on the very kinds of aerial analytics workloads embodied in our own repository and in systems like SpatialSky. 

#codingexercise: CodingExercise-01-07-2026.docx