There's a quiet worry underneath a lot of agent-assisted work. If the agent writes the code and we only check it, do we stay good enough to check it? We don't have a tidy answer. We do have two sources worth reading side by side, one very old and one very new.
The irony from 1983
Lisanne Bainbridge wrote about industrial plants, not software. Her paper, "Ironies of Automation", argued that automating most of a process leaves a person with the job of watching it and stepping in when it fails, and that this is harder than it sounds.1
One of her points is about practice. She notes that a formerly experienced operator who has been monitoring an automated process may now be an inexperienced one, because skills fade when they aren't used.1 The cruel part is the timing. The moments an operator is needed are the unusual ones, and those are the moments when rusty skills cost the most.
A plant control room isn't a terminal with an agent in it, and we wouldn't push the comparison far. Still, the shape of the worry travels well. Supervising work you no longer do often is a skill of its own, and it needs the same upkeep as the doing.
A 2026 trial that looks a bit like it
In January 2026, Anthropic researchers published a randomized trial of how AI assistance affects learning. The study recruited 52 mostly junior software engineers who were unfamiliar with Trio, a Python library, and had them complete tasks with it.2
Afterwards a quiz checked what they'd absorbed. According to the write-up, the AI group averaged 50% on the quiz, compared to 67% in the hand-coding group. The largest gap was on debugging questions.2 The AI group finished about two minutes faster, but the difference wasn't statistically significant.2
Debugging is exactly the skill Bainbridge's operator needs on a bad day, so that detail stuck with us. It's one small study, though, with one library and a quiz taken right after the task. It doesn't tell us what happens over months, or what happens to you.
It depended on how people used the assistant
The result wasn't uniform. The authors describe several interaction patterns, and the highest scorers, those averaging 65% or more on the quiz, tended to stay engaged with the material. Some asked the assistant follow-up questions to improve understanding after generating code. Others asked for code along with explanations. A third group asked only conceptual questions and relied on their own understanding to finish the task.2
The groups were small, so we'd hold the labels loosely. But the direction is gentle and useful. The difference wasn't whether the assistant was there. It was whether the person was still doing some thinking.
What we take from the pair
Neither source says agents make people worse at their jobs. Bainbridge wasn't writing about software at all, and the trial measured learning something new, not maintaining skills you already have. What they share is a caution: understanding is something we practise, and it doesn't stay put just because we once had it.
That's a modest claim, and we think it's the right size. It means the question to ask of a session isn't "did the agent do it?" but "could I explain it, and could I fix it if it broke?"
Where Cartographer fits
Cartographer reads the session history your coding agent already keeps and turns it into a map of how a project was built. It runs on your machine, with no cloud. We mention it only because a record of what happened is easier to learn from than a memory of it, and the habit below goes better with something to look back at.
A small habit
Pick one change from the week that the agent mostly wrote. Without looking at the diff, try to explain in a few sentences what it does and where you'd look first if it broke. Then check. If you stumble, ask the agent to explain the part you missed, or trace it by hand. It takes ten minutes, and it keeps the supervising muscle in use. If it goes well, you've lost nothing. If it doesn't, you've found out cheaply, on a quiet afternoon, instead of during an outage.
Sources
- Bainbridge, L. Ironies of Automation. Automatica, 19(6), 775-779, 1983.
- Shen, J. H., Tamkin, A. How AI assistance impacts the formation of coding skills. Anthropic research, January 29th 2026 (arXiv title: "How AI Impacts Skill Formation").