- Published on
Two Modes of Agentic AI Work
- Authors
- Name
- DeepHorus
In a recent essay titled "The Greenhouse and the Lens: Two Modes of Agentic AI Work," an author argues that AI coding assistants can be used in two distinct ways, and mistaking one for the other leads to frustration. The greenhouse traps ambient heat to nurture many plants, while a lens concentrates light to ignite a single point. These metaphors describe how developers interact with tools like Claude Code.
The author initially believed the modes corresponded to junior and senior engineers. However, they found the mapping false: neither mode is inherently advanced. Instead, they are distinct tools for distinct jobs, and much dissatisfaction with agentic AI comes from using the wrong mode.
In greenhouse mode, the user doesn't know exactly what they want and lets the AI generate many outputs to be evaluated later. The author's AxiDraw plotter experiments are cited as a pure example, where Claude Code produced dozens of plotter-art tools. Most were mediocre, a few delightful, and none were planned in advance.
Similarly, a small macOS/SwiftUI app that converts Markdown to Word documents was built without a spec or design. The author notes the cost of trying things out has dropped to nearly zero, making this exploratory approach viable.
Lens mode is the opposite discipline: the user knows exactly what "done" looks like and concentrates all effort on reaching it. The agent proposes ten things, and nine are discarded because they aren't on the path. This mode offers leverage on a known target, but the target must be chosen by the user.
The author asserts that the truly advanced skill is recognizing which mode the current work requires and shifting when drift occurs. An example is sitting down to ship a specific feature but ending up letting the AI scaffold an admin panel—a greenhouse activity in what should be lens work.
The essay connects this to a previous argument that domain expertise is the real moat. In both modes, a verified mental model serves as an oracle to pick the right ideas or reject plausible-but-wrong outputs. As agentic AI grows more capable, users must develop the meta-skill of asking "greenhouse or lens?" to keep their work on track.
