--- title: Agentic Skill Decay | daily.dev description: Addy Osmani argues that relying on coding agents can short-circuit the trial-and-error process that historically built engineering expertise. He cites an... canonical: https://daily.dev/posts/agentic-skill-decay-nzqkcham8 twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Agentic Skill Decay | daily.dev og:description: Addy Osmani argues that relying on coding agents can short-circuit the trial-and-error process that historically built engineering expertise. He cites an... og:url: https://daily.dev/posts/agentic-skill-decay-nzqkcham8 og:image: https://api.daily.dev/og/posts/nZQKcHaM8.png og:image:alt: Agentic Skill Decay og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Agentic Skill Decay **[Addy Osmani](https://daily.dev/sources/addy)** · [@addyosmani](https://daily.dev/addyosmani) · 19 min read · 81 upvotes · 13 comments ## Summary Addy Osmani argues that relying on coding agents can short-circuit the trial-and-error process that historically built engineering expertise. He cites an Anthropic study where junior engineers using AI assistants scored 50% on a follow-up quiz about a Python library versus 67% for those working by hand, with better outcomes tied to asking conceptual questions rather than just prompting for code. He also references an Anthropic study of ~400,000 Claude Code sessions showing that even intermediate task-specific expertise improves the odds of reaching verified success. His advice: deliberately form hypotheses before prompting, ask agents to explain their work, capture lessons into repo-level artifacts (lint rules, tests, docs) rather than relying on chat memory, and treat verification skill and imagination as the real bottlenecks now that agents can execute almost anything. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### Does using AI coding assistants hurt learning compared to writing code by hand? Yes, according to a 2026 Anthropic study of junior engineers learning the Python library Trio: those who used AI assistants scored 50% on a follow-up quiz, compared to 67% for those who worked by hand. Within the AI-assisted group, the strongest results came from people who asked conceptual questions and requested explanations rather than treating the model purely as a code generator. _Developers weighing AI-assisted learning tradeoffs can follow deeper coverage of this research on daily.dev._ ### Does having domain expertise actually improve success rates when using Claude Code? Yes, an Anthropic study of around 400,000 Claude Code sessions found that even intermediate expertise in the task domain increased the likelihood of reaching verified success, compared to more novice users. Deep decade-long experience wasn't required, but understanding the problem domain well enough to recognize what a good outcome looks like made a measurable difference. _Anyone judging how much domain knowledge still matters with agentic tools can track findings like this on daily.dev._ ## Community discussion Top comments from developers on daily.dev. **@iaparatudia** · 9 upvotes > This is becoming one of the biggest questions for me as I use coding agents more: how do I increase execution speed without outsourcing the learning? > > One practice I'm starting to value is turning what I learn with an agent into durable project artifacts — tests, docs, rules, skills — instead of leaving the knowledge buried in a chat. > > The agent can accelerate the implementation. I still need to understand why the implementation works and develop the judgment to know when it doesn't. **@ahmetozel** · 9 upvotes > The 50 versus 67 result is the number to sit with, and the qualifier matters: the gap tracked whether people asked conceptual questions rather than just requesting code. So it is not that AI makes you worse, it is that delegating the framing makes you worse. Capturing lessons into lint rules, tests and docs rather than chat memory is the advice I would put first of the four. Anything living only in a session is gone next week, and a rule in the repo is the one form of learning that also constrains the agent next time. Forming a hypothesis before prompting is the scientific method retrofitted... **@akkitto** · 5 upvotes > Trial & Error - This is a short summary of my whole life. **@confidentcoding** · 5 upvotes > _"Most of the judgment I use today came from thousands of small reps like these: debugging failures, reviewing other people’s code, and living with abstractions that looked good until a real system pushed back. Agents can now skip much of that work."_ > > I've been thinking of a good analogy for this. I know in The Pragmatic Programmer, the author refers to software development as more akin to "gardening" then any other field like construction or accounting. The reason is because in gardening you always have to make a lot of micro decisions. You often pivot based on new information that you're... **@johhnypav** · 3 upvotes > On top of all this, the AI coding agents/assistants need to have a proper organizational context to work efficiently. So they understand everything and actually go in-depth instead of just hallucinating. --- Tags: [#ai-coding](https://daily.dev/tags/ai-coding), [#claude-code](https://daily.dev/tags/claude-code), [#code-review](https://daily.dev/tags/code-review) [View this post on daily.dev](https://daily.dev/posts/agentic-skill-decay-nzqkcham8) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Agentic Skill Decay","url":"https://daily.dev/posts/agentic-skill-decay-nzqkcham8","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/agentic-skill-decay-nzqkcham8"},"datePublished":"2026-09-20T18:35:04.149Z","dateModified":"2026-09-21T15:00:13.347Z","description":"Addy Osmani argues that relying on coding agents can short-circuit the trial-and-error process that historically built engineering expertise. He cites an...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/ea2545340d64b8d4ec6bb32838907ad5?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/ea2545340d64b8d4ec6bb32838907ad5?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Addy Osmani","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Person","name":"Addy Osmani","url":"https://daily.dev/addyosmani","image":"https://avatars2.githubusercontent.com/u/110953?v=4","interactionStatistic":{"@type":"InteractionCounter","interactionType":{"@type":"EndorseAction"},"userInteractionCount":5510}},"commentCount":13,"discussionUrl":"https://daily.dev/posts/agentic-skill-decay-nzqkcham8","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":81},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":13}],"keywords":"ai-coding,claude-code,code-review","timeRequired":"PT19M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Addy Osmani","item":"https://daily.dev/sources/addy"},{"@type":"ListItem","position":3,"name":"Agentic Skill Decay"}]} {"@context":"https://schema.org","@type":"WebPage","@id":"https://daily.dev/posts/agentic-skill-decay-nzqkcham8","comment":[{"@type":"Comment","text":"This is becoming one of the biggest questions for me as I use coding agents more: how do I increase execution speed without outsourcing the learning?\nOne practice I’m starting to value is turning what I learn with an agent into durable project artifacts — tests, docs, rules, skills — instead of leaving the knowledge buried in a chat.\nThe agent can accelerate the implementation. I still need to understand why the implementation works and develop the judgment to know when it doesn’t.","datePublished":"2026-09-20T22:27:46.195Z","url":"https://daily.dev/posts/nZQKcHaM8#c-2nH0hdvxU","author":{"@type":"Person","name":"Julius AQ","url":"https://daily.dev/iaparatudia","image":"https://avatars.githubusercontent.com/u/49326430?v=4"},"interactionStatistic":{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":9}},{"@type":"Comment","text":"The 50 versus 67 result is the number to sit with, and the qualifier matters: the gap tracked whether people asked conceptual questions rather than just requesting code. So it is not that AI makes you worse, it is that delegating the framing makes you worse. Capturing lessons into lint rules, tests and docs rather than chat memory is the advice I would put first of the four. Anything living only in a session is gone next week, and a rule in the repo is the one form of learning that also constrains the agent next time. Forming a hypothesis before prompting is the scientific method retrofitted onto tooling, which is slightly embarrassing to have to rediscover but clearly works.","datePublished":"2026-09-20T20:36:44.329Z","url":"https://daily.dev/posts/nZQKcHaM8#c-gFrC9FDOC","author":{"@type":"Person","name":"Ahmet Özel","url":"https://daily.dev/ahmetozel","image":"https://avatars.githubusercontent.com/u/70992231?v=4"},"interactionStatistic":{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":9}},{"@type":"Comment","text":"Trial & Error - This is a short summary of my whole life.","datePublished":"2026-09-21T19:58:43.658Z","url":"https://daily.dev/posts/nZQKcHaM8#c-68D8yVnjk","author":{"@type":"Person","name":"Daniel","url":"https://daily.dev/akkitto","image":"https://media.daily.dev/image/upload/s--FtwJqX4c--/f_auto/v1754900041/avatars/avatar_29TCpY2hJR72V3BlxPXzX?_a=BAMClqZW0"},"interactionStatistic":{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":5}},{"@type":"Comment","text":"“Most of the judgment I use today came from thousands of small reps like these: debugging failures, reviewing other people’s code, and living with abstractions that looked good until a real system pushed back. Agents can now skip much of that work.”\nI’ve been thinking of a good analogy for this. I know in The Pragmatic Programmer, the author refers to software development as more akin to “gardening” then any other field like construction or accounting. The reason is because in gardening you always have to make a lot of micro decisions. You often pivot based on new information that you’re getting that you didn’t previously have.\nIf we continue to use this analogy: using coding agents is like having a landscaper come and plant an entire garden for you, but you have never even watered a house plant or picked up a trowel.\nYou took the shortcut and it initially seems awesome, but you have no idea what to do or how to maintain it…and you have to play a LOT of catch up on your learning if you want to make sure the garden doesn’t fall into chaos or completely die.\nYou skipped the fundamentals so you could receive the most high: an end product that normally people work hard at. They don’t work hard at it just to get to the end goal, they also gain all the knowledge skills needed to support it long term. Especially if they ever want to plant another garden!","datePublished":"2026-09-21T19:35:57.897Z","dateModified":"2026-09-21T19:38:58.915Z","url":"https://daily.dev/posts/nZQKcHaM8#c-f5vqFIm1N","author":{"@type":"Person","name":"Lars Faye | Confident Coding","url":"https://daily.dev/confidentcoding","image":"https://media.daily.dev/image/upload/s--OGZu5DEc--/f_auto/v1772569630/avatars/avatar_umWZ9aQAng34qk5aaJl2q?_a=BAMAMiiu0"},"interactionStatistic":{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":5}},{"@type":"Comment","text":"On top of all this, the AI coding agents/assistants need to have a proper organizational context to work efficiently. So they understand everything and actually go in-depth instead of just hallucinating.","datePublished":"2026-09-21T07:02:55.831Z","url":"https://daily.dev/posts/nZQKcHaM8#c-UeNcQqqFH","author":{"@type":"Person","name":"Pavan Belagatti","url":"https://daily.dev/johhnypav","image":"https://media.daily.dev/image/upload/s--r9dshbly--/f_auto/v1776930121/avatars/avatar_DFA0U7-fS?_a=BAMAMiWQ0"},"interactionStatistic":{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":3}}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/agentic-skill-decay-nzqkcham8#faq","mainEntity":[{"@type":"Question","name":"Does using AI coding assistants hurt learning compared to writing code by hand?","acceptedAnswer":{"@type":"Answer","text":"Yes, according to a 2026 Anthropic study of junior engineers learning the Python library Trio: those who used AI assistants scored 50% on a follow-up quiz, compared to 67% for those who worked by hand. Within the AI-assisted group, the strongest results came from people who asked conceptual questions and requested explanations rather than treating the model purely as a code generator. Developers weighing AI-assisted learning tradeoffs can follow deeper coverage of this research on daily.dev."}},{"@type":"Question","name":"Does having domain expertise actually improve success rates when using Claude Code?","acceptedAnswer":{"@type":"Answer","text":"Yes, an Anthropic study of around 400,000 Claude Code sessions found that even intermediate expertise in the task domain increased the likelihood of reaching verified success, compared to more novice users. Deep decade-long experience wasn't required, but understanding the problem domain well enough to recognize what a good outcome looks like made a measurable difference. Anyone judging how much domain knowledge still matters with agentic tools can track findings like this on daily.dev."}}]} ```