--- title: "AI skill atrophy: why agent operators should keep one unaided rep" slug: harness-ai-skill-atrophy-unaided-reps date: 2026-10-26 excerpt: Experienced doctors who worked with an AI assistant for a few months found fewer polyps once it was switched off. The same quiet loss can happen to anyone who reviews AI agents all day, and the fix is a small, regular piece of work done without the agent. featured_image: "https://bbtxujdxvidaghmhxkqs.supabase.co/storage/v1/object/public/generated-images/blog-1792994733093-harness-ai-skill-atrophy-unaided-reps.webp" featured_image_alt: "Illustration of a workbench where a person sketches a plan by hand on paper while a small screen beside them shows a read-only magnifying-glass icon." author: Kinan Hamwi author_url: https://www.linkedin.com/in/kinanhamwi/ canonical_url: https://cerevisor.com/blog/harness-ai-skill-atrophy-unaided-reps updated_at: 2026-10-26T06:05:34.646991+00:00 --- # AI skill atrophy: why agent operators should keep one unaided rep By the founder of Cerevisor. Checked on 26 October 2026 against the source code of [Cerevisor 3.9](/blog/cerevisor-v39-agent-groups-free-canvas). TLDR A study of experienced doctors found their unaided detection rate fell after a few months of working with an AI assistant. Anyone who spends the day checking AI agents can lose skill the same quiet way. The cheapest protection is one small, regular piece of work done by hand, with an agent that can only read and comment. This week a short write-up went around about doctors and polyps. It retold a study from The Lancet Gastroenterology & Hepatology: experienced endoscopists, the doctors who look for early signs of bowel cancer during a colonoscopy, used an AI assistant for several months. When they went back to working without it, they found fewer of the growths they were looking for. I run a team of AI agents most days, so I read it as a note about myself. --- ## The myth that review keeps your skills sharp The comfortable belief goes like this: if I check every piece of [agent work](/blog/harness-ai-agent-governance-patterns), I stay in the loop, so my own skill stays where it was. Reading code is still engineering. Approving a plan is still planning. It sounds right because review really is hard work. It is tiring, it takes judgment, and it catches real mistakes. Effort feels like practice. ## What the colonoscopy study measured The study did not ask doctors how they felt. It counted results. Skill atrophy here means a plain thing: a skill that gets weaker because it is used less. > "Before AI was introduced into their workflow, their adenoma detection rate in standard procedures stood at 28.4%. After several months of routine AI-assisted work, that figure fell to 22.4% when they returned to non-AI procedures, a relative decline of about 20%." Quest Review Center, October 2026, reporting the Lancet Gastroenterology & Hepatology study These were not beginners. The write-up says the study tracked 19 doctors at four Polish centres, each with more than 2,000 colonoscopies behind them. The authors were careful: the study observed what happened and cannot prove the AI alone caused the drop. I hold it the same way. It is a strong warning, not a law. 28.4% to 22.4% unaided detection rate before and after months of AI-assisted work, in the reported study The useful part is the shape of the loss. Nobody lost a job. The work still got done every day. What faded was the part of the skill the machine had been doing for them, and it only showed when the machine was switched off. ## Why reviewing AI agent work is not the same practice Here is the gap in the myth. Checking an answer and producing one use different muscles. When an agent writes the plan, the human never has to hold the whole problem in their head, pick a starting point, or get stuck. Getting stuck is where a lot of skill is built. The same write-up points to aviation, where the US Federal Aviation Administration has long recommended building manual flying back into routine operations, so pilots keep the skill they need on the day the autopilot gives up. For someone running agents, the day the autopilot gives up looks ordinary. A provider is down, a model behaves oddly after an update, or an agent keeps proposing a fix that is subtly wrong. That is the moment the operator’s own unaided judgment is the only thing left, and it is the one skill that review alone does not exercise. > Review keeps you able to spot a wrong answer. It does not keep you able to produce a right one. --- ## Unaided practice with AI agents: flip who does the work The answer is not to stop using agents. It is to keep a small number of repetitions for yourself, on purpose, and to let an agent check you instead of the other way round. In Cerevisor this is a setting, not a discipline you have to remember. Every agent has an Agent Type, which decides the tools it starts with. The Explore type is read-only: it can read, search and list the files in a workflow’s folder, and find skills, but it cannot write or edit a file or run a command. The [Agent roles reference](https://cerevisor.com/docs/reference/agent-roles) lists what each type can do. So the weekly routine I use looks like this. I pick one small task I would normally hand off, such as tracing why a test fails or sketching the steps of a change. I do it by hand. Then I give an Explore agent the folder and one question: what did I miss? Because it cannot touch the files, it cannot quietly finish the work for me. It can only point. Key Insight A read-only agent turns the usual arrangement around. The person produces, the agent checks, and the skill that fades fastest gets its repetition back. ## A read-only AI agent is a rule the app enforces The same read-only rule already protects ordinary reviews in Cerevisor. When an agent is set to review another agent’s work, Cerevisor’s workflow check reports an error if that reviewer has been given tools that can write or edit, or if its Agent Type is anything other than Explore. The reviewer is meant to judge, not to change what it is judging. I find that reassuring for the flipped routine too. The boundary between doing and checking is not a promise in a prompt. It is a list of tools the agent simply does not have. ## Keep one unaided rep this week Pick one task you have handed to an agent at least three times this month. Do it yourself once, then open the [agent settings guide](https://cerevisor.com/docs/guides/building-workflows/add-and-configure-agents), set one agent’s type to Explore, and ask it what you got wrong. Notice where you hesitated. That hesitation is the skill telling you where it needs the rep. #### Sources - [AI-assisted doctors miss more polyps when the machine is switched off](https://questreviewcenter.com/software-reviews/ai-assisted-doctors-miss-more-polyps-when-the-machine-is-switched-off/) - Quest Review Center, 2026-10-25 - [Agent roles](https://cerevisor.com/docs/reference/agent-roles) - Cerevisor Docs - [Add and configure agents](https://cerevisor.com/docs/guides/building-workflows/add-and-configure-agents) - Cerevisor Docs