The plant is not the unit

Agtech's favorite promise is that we will soon farm plant by plant. I think that promise is aimed at the wrong unit.

The vision is easy to love. Cameras on a sprayer boom can already tell a weed from a crop seedling in milliseconds. Robots can thin, weed and sample individual plants. Add enough sensing and compute, the story goes, and every plant gets its own dose of nitrogen and water.

It sounds like the end point of precision agriculture. I think it runs against what evolution teaches about crops. A maize field is a population, and its yield belongs to the population, which is a different thing from adding up what each plant could do on its own.

I learned this, without quite noticing, in a lettuce field.

What natural selection optimizes

Natural selection rewards the individual that leaves more offspring than its neighbors. A farmer wants the group that produces the most in total. In a crop, those two goals pull apart.

R. Ford Denison built a whole book on this point, Darwinian Agriculture (2012). Wild plants evolved to win against their neighbors: grow a little taller to shade them, spread leaves flat to intercept their light, build more roots to grab nutrients first. Each of these traits helps the plant that has it and costs the stand. Ecologists call this a tragedy of the commons, and it plays out in every field.

Roots show it clearly. In a glasshouse experiment, soybean plants that shared soil with a neighbor built 85% more roots than plants with soil to themselves, while the plants with their own space produced about 30% more seed (Gersani et al., 2001). Those extra roots went into the fight with the neighbor, and the seed paid for them. The design is still debated, because sharing also changes how much soil each plant gets, but I find the pattern hard to ignore.

Denison draws a practical conclusion from this. Traits that natural selection has already refined for millions of years, like the basic efficiency of photosynthesis, leave little room for us to improve. Traits where individual interest and group interest conflict, like height and leaf angle, leave a lot.

Victor Sadras, an Argentine crop physiologist now in Australia, has measured the trade-off directly. With Mariano Cossani he grew wheat cultivars released over five decades in pairs. Yield gains tracked a loss of competitive ability: newer cultivars competed less for light and used it more efficiently (Cossani and Sadras, 2021). Denison put the lesson plainly: much of the yield potential breeders gained came from reversing past selection for individual-plant competitiveness (Denison, 2015).

We already proved it once

The biggest yield jump of the last century came from making plants worse at competing. Nobody called it that at the time.

The Green Revolution's semi-dwarf wheat and rice were short, stiff plants that put less into stems and more into grain. Alone, a short plant loses to a tall neighbor that shades it. In a dense, well-fed field of equally short plants, nobody wastes carbon racing for light, and the stand stays upright under heavy heads.

In 1968 the Australian agronomist Colin Donald proposed breeding exactly this on purpose. His wheat "ideotype" was a weak competitor: one stem, few small upright leaves, a modest root system. He wanted a plant that did well among copies of itself, even if it would lose a one-on-one contest with a wild neighbor.

At the University of Copenhagen, Jacob Weiner and colleagues turned Donald's idea into a research program they call evolutionary agroecology. Natural selection taught each plant a defensive move under competition: shade avoidance, stretching to escape its neighbors. Weiner proposed the opposite for crops, an offensive, collective "shading" strategy in which dense, evenly sown cereals use their head start to shade weeds out, and argued that breeders should select groups rather than individuals to get there (Weiner et al., 2010). In their Danish wheat trials, raising crop density cut weed biomass and raised crop biomass every year (Olsen et al., 2012).

The idea has held up in yield trials too. Weiner and colleagues grew 35 wheat varieties both as single-variety plots and as a mixture of all of them. The varieties that competed best in the mixture gave mediocre yields in their own plots; the less competitive ones yielded more (Weiner et al., 2017). A later study added a nuance that matters for farming: the conflict shows up when water and nutrients are plentiful, and fades when they are scarce (Yang et al., 2024).

Animals show the same pattern. William Muir at Purdue selected laying hens as whole cages rather than as individual top layers, which tends to reward hens that succeed at their cage-mates' expense. In six generations, annual mortality in the group-selected line fell from 68% to under 9%, and eggs per hen housed rose from 91 to 237 (Muir, 1996).

Why plant-by-plant management repeats the mistake

Managing each crop plant toward its own optimum treats the plant as the unit of yield. Breeders spent a century learning otherwise.

What a plant "needs" depends on its neighbors, because they are drawing on the same light, water and nitrogen. Feed the laggard and it shades the plant next to it. Push one plant to its maximum and you often pay for it with the two beside it. Per plant, it looks like an improvement; per hectare, it can be zero or worse.

Stands also correct themselves. Wheat fills gaps by tillering, soybean by branching. A stand absorbs a lot of plant-to-plant variation without losing yield, which is why rescuing individual plants rarely pays. Much of what plant-level sensing shows us is variation the population already handles.

Sunflowers show this better than anything I know. Mónica López Pereira, working with Antonio Hall, Victor Sadras, Jorge Casal and William Batista at the University of Buenos Aires, found that neighboring sunflowers in dense rows lean in opposite directions. The zigzag starts with a few pioneer plants and spreads down the row as each plant reads the red to far-red light reflected by its neighbors. Stands left to organize themselves produced 19% to 47% more oil than stands forced to stay upright (López Pereira et al., 2017).

Nobody writing prescriptions plant by plant would have designed that zigzag. An engineer optimizing each plant would have kept them all upright, the ideal posture for a plant alone, and lost a fifth to almost half of the oil. The plants found the better arrangement by reacting to each other.

Plant-level sensing is still useful. I see it as an input to decisions made at a larger scale. Telling a weed from a crop plant is a real plant-level problem, because the weed belongs to a different population. Stand counts, early disease foci and emergence gaps are plant-level data that feed population-level choices.

The industry's marketing often says the opposite. When John Deere's See & Spray was honored at CES 2022, the company called it plant-level management, and its CTO said the goal was to "treat each plant individually" (Deere, 2021).

The technology underneath started somewhere else. Before Deere bought Blue River Technology in 2017, Blue River's first product was LettuceBot, a thinning machine. Lettuce growers sow more seed than they need. LettuceBot's cameras looked at every seedling and sprayed the extras, keeping the best plants at the right spacing, more than 5,000 decisions a minute (Vegetable Growers News, 2017).

It looked at plants one at a time, but what it managed was the stand: how many plants, how far apart, how even. I worked there, and I still think that original framing was right.

The decisions that move yield sit at other scales. Density, row spacing and uniformity are decisions about the stand, and Weiner's work shows that dense, evenly spaced cereals suppress weeds because the crop plants end up competing with the weeds more than with each other. Choosing cultivars that perform well as a group is a breeding decision, and intercrops and cultivar blends are decisions about the mixture. A crop's job also continues after harvest, because it sets up the soil, pests and diseases for whatever comes next. I wrote more about that last one in The crop after.

Technology at the right scale

I am not against robots, sensors or satellites. I care about which unit they serve.

I have worked on both ends of this. At Blue River we built machines that saw every lettuce plant, yet the decision they made was about the stand. At Planet we look at fields from orbit, one stand and one season at a time. Both jobs taught me that data about single plants is most useful when it feeds decisions about the whole stand.

Donald, Denison, Sadras, Weiner, Hall and López Pereira worked these ideas out years ago; I am only passing them along. I want more people building agtech to know their work. A plant optimized for itself is what natural selection already produced, and the Green Revolution's yield gains came from moving away from it.