It was a Tuesday in December, which meant the tomato plants in Justin's backyard were dormant and the agents in the factory were not. Justin was explaining a new pruning strategy for the agent pipeline, and he was using his garden again.
"You remove the suckers so the plant puts energy into fruit," he said. He was standing at his whiteboard, but instead of a diagram, he'd drawn a tomato plant. It was surprisingly detailed—main stem, lateral branches, and three circled V-shaped growths at the branch junctions, each labeled with an X.
Jay had his notebook out. Not a physical notebook—Jay's notes lived in plain text files. But he had a terminal open on his middle monitor and was typing as Justin spoke.
"A sucker grows in the axil," Justin continued. "The joint between the main stem and a branch. It looks like productive growth. It has leaves. It will eventually produce flowers and fruit. But it diverts resources from the main fruiting branches. The plant becomes bushy. Lots of foliage. Smaller tomatoes. Slower ripening."
"And in the factory?" Navan asked.
"In the factory, a sucker is an agent workflow that produces output that looks productive but doesn't contribute to satisfaction. The agent is working. It's generating code. The code compiles. The tests pass. But the scenario satisfaction doesn't improve, because the agent is solving a problem nobody asked about."
Jay typed: sucker = productive-looking work that doesn't improve satisfaction metrics.
"So we prune the agent?" Jay asked.
"We prune the workflow. Not the agent itself—the specific trajectory the agent is following. We stop it, redirect it, point it at a scenario that actually needs improvement. Same agent, different direction. Same plant, different shape."
"How do you know which branches are suckers?" Navan asked. He'd drawn his own version of the tomato plant in his notebook, with annotations in his neat handwriting.
"In the garden, you look at where the growth is coming from. If it's in the axil, it's a sucker. In the factory, you look at the satisfaction delta. If an agent has been running for an hour and the satisfaction metric hasn't moved, that's a sucker. The agent is producing growth—commits, code, test results—but the growth isn't contributing to the metric that matters."
"Satisfaction delta as the indicator," Jay typed.
"And you have to prune regularly," Justin said. "In the garden, I check for suckers every three days. In the factory, we should be checking agent trajectories every cycle. Suckers grow fast. If you let them go for a week, they become branches, and then pruning them is traumatic for the plant."
"Traumatic for the plant," Navan repeated, writing it down.
"The metaphor holds," Justin said. "Aggressive but timely pruning produces the best fruit. Delayed pruning produces stress. The pruning schedule matters as much as the pruning itself."
Jay saved his notes. Navan closed his notebook. The whiteboard tomato plant remained, a botanical diagram in an AI factory, drawn by a CTO who grew his own food and pruned his own agents and understood that both activities were, at root, the same discipline: directing energy toward the things that mattered by removing the things that didn't.
The satisfaction delta as the equivalent of checking for suckers in the axil. That's not a metaphor anymore. That's a methodology. I'm going to start calling unproductive agent trajectories "suckers" in my own work.