Shane Thomas's recent piece in Upstream Ag Insights, "The Evolution of Spray Technology and the Tech Stack" (June 2026, subscription required), is one of the clearest frameworks I've seen for where precision crop protection is actually heading. Layering software-enabled decision support, camera-augmented application, and droplet enhancement into a shared mental model helps explain why no single tool solves the problem, and why combinations of capabilities matter. It gives the industry a common language that it has needed for a while.
From where I sit at Verdant Robotics™, that framing is right. And it surfaces the next problem.
As each layer of the stack improves, the system becomes more capable and more fragmented at the same time. Decision tools get better at telling growers when and where to act. Vision systems get better at identifying weeds. Hardware gets better at placing product. But the layers stay loosely connected. Insights generated in one system get executed in another. Feedback arrives hours or days after application. Optimization happens inside each layer rather than across the whole system.
We are building a better stack. We are not yet building a unified one.
Shane's middle layer, precision spraying and camera-augmented application, is where this tension is sharpest. The dominant architecture in that layer mounts vision systems onto existing sprayer hardware. The camera sees. A separate system decides. The nozzle applies. Three distinct components, often from different vendors, communicating across handoffs. Each handoff is a place where latency, translation loss, and optimization gaps accumulate. That is not a criticism of the companies building in that layer. It is a structural observation about the architecture itself.
When we started Verdant, we made a deliberate choice not to follow that path. Rather than adding a better camera layer to an existing sprayer, we designed the perception-to-decision-to-execution loop as a single native system at the plant level. This is physical AI applied to the field: embodied intelligence that doesn't just analyze data but directly controls physical action in real time. What makes that possible is spatial AI at the plant level. The system builds a real-time spatial model of every plant in its path, understanding crop geometry, canopy structure, and target position in three dimensions. That spatial model is what powers Aim & Apply™, our closed-loop technology that detects and tracks each target, predicts where it will be at the moment of application, and delivers a micro-liter shot on target. SharpShooter™ is the expression of that architecture.
In practice, that comes down to three design choices we made early and have held to.
Hardware and perception are co-designed. Cameras, lighting, and mechanical systems are built to work together from the start, enabling the system to construct accurate spatial models of crops and weeds at millimeter-level precision in dense canopies and tight row spacing. That is why we can operate in crops like grass seed and sod, where the crop and the weed can look nearly identical to the human eye. Camera-augmented application was not designed for that environment. A native spatial AI system was.
Autonomy is the default, not a layer on top. The on-board AI does not produce suggestions for a human to interpret or recommendations for a separate system to execute. At field speed, it builds a spatial model of what is in front of it, identifies individual targets, and controls application turrets in real time. Aim & Apply™ closes that loop continuously, adjusting timing and placement to maintain millimeter-level accuracy as speed, density, and field conditions change. The system aims before it applies rather than broadcasting and adjusting after the fact.
The system is built to learn from every pass. This is where Hindsight comes in. Hindsight continuously captures what the system sees and does in the field, every plant, every shot, and feeds that back into the autonomy stack. Over time, the spatial models improve. Detection gets sharper. Shot placement gets tighter. Perception, action, and learning are part of the same loop. Each treated acre makes the system more accurate on the next one. That is not a data feature. It is a compounding structural advantage. A loosely coupled stack cannot replicate it because there is no single system closing the loop.
Using Shane's framework directly: Verdant compresses the decision layer and the precision spraying layer into one orchestrated system. The software-enabled decision layer becomes an on-board autonomy stack that ingests sensor data and acts immediately, with no handoff to separate hardware. The camera-augmented application layer is replaced by a fused system where the same spatial AI that builds the plant model controls the turrets that apply. Execution optimization is not a separate problem. It is part of a broader control question: what is the right intervention for this plant, at this moment, given both agronomic and economic objectives?
That design shift changes the core question. Much of the current stack, as Shane describes it, is focused on improving application mechanics: reducing drift, improving coverage, optimizing gallons per acre. Those are the right problems to work on within the existing architecture. But physical AI grounded in spatial reasoning is designed around a different objective: biological outcomes per plant. The question stops being "how do we make this spray event more efficient?" and becomes "how do we deliver the right intervention, at the right intensity, to the right plant, at the right time?"
For growers, that shows up as fewer passes, lower input costs, and more consistent results. For investors, it shows up as durable unit economics in high-value, high-regulation crop systems.
Shane's framing moved the conversation past "this tool versus that tool" toward "how do these capabilities stack to deliver better agronomy, economics, and environmental outcomes?" That is the right question.
The one I would add is: when does it make more sense to rethink the architecture than to keep adding layers to it?
Native, plant-level systems that close the loop through learning are not a replacement for the stack. They are what happens when the stack matures into something unified.
