Coding with agents is a trap, and we all fell for it.

The author of the article left this comment, wanted to share here

You say something along the lines of “knowing parts of your codebase in depth is important” How deep are we talking about here - 1. Knowing my codebase uses react 2. Knowing jsx, useState, useEffect etc. 3. Knowing how diffing algorithm and react fiber works under the hood.

3 isn’t your codebase. 1 and 2 are basic knowledge I expect you to have. I’m mostly talking about the architecture and relationship between parts. You should be able to map out how data flows in your app when users interact.

@HiteshRohira15 Most engineers can’t do this for a code base of any size or set of services making up a system. It’s rare when you can find someone who does.

@theo > Most engineers can’t do this I am surprised to hear this, I would have thought this was the bare minimum?

The skill atrophy from not coding can get pretty serious, but at least many of us spent years in the industry before AI. So even if the syntax gets rusty, the principles and the logical and analytical thinking we built up will still be there. The real issue will be future devs and software engineers entering the field without ever having written a single line of code… they’re cooked.

Well I wrote something on it just yesterday, seems similar but im not good in articulating my thoughts so just putting it out here:

LACK works great, idk.

Stream of facts (Me agreeing with a guy agreeing with a guy)

I forced myself to build the foundations of my new application without AI, just to see if I still had the same willingness to solve problems that would occur, reading the documentation myself. It is a good test that will make you reflect a lot on how your work habits have changed

To me agentic coding isn’t that much different than what you end up doing as a senior+ engineer anyway. Most of my days even before coding agents were strategy, breaking down ambiguity into something concrete, and reviewing code, etc. The sheer volume is what makes it hard

I've been saying this for years (or however long agents have been out, it sure feels like years). Agents have a number of problems that Anthropic/OpenAI will never address because profit

In case someone is looking for the original Article (which I couldn't find here, but was in Theo's YT video description):

The trap isn't agents. It's mistaking output for ownership.

I love your videos and I've been watching them for years, but from that video I've gotten a lil' sad I felt like you were saying "I can use llms, I'm smart, you can't, you're dumb", but probably that's just my ego talking. That made me realize there's too much nuance to this topic to put it into one video I'd love to have a conversation about this someday - I'm a bit inbetween the generations - when ChatGPT first released, I was in my 1st semester of Master's degree - and when AI first became really useful (I'd say around supermaven) I was finishing my master's degree, and by that time I already had 4 years of work experience I think it's not that "all older devs are better and younger ones are stupid" - it's just that statistically some very small amount of devs are going to use the tools in a "correct way" whatever that means - just like you said you know 17-year olds that push the industry forward notwithstanding their age It's interesting to me that you didn't touch the subject of people who hate AI. Maybe it's just people I know, but really I know quite a few people who just refuse to do anything with AI because it's "evil" and I believe they'll have it just as bad as the one's who mindlessly ship slop without having at least a mental map of any given project

Never seen a Theo vid get this much hate and it was like pretty tame like it was just going an article

I didn’t see much hate on this one?

Idiots on YouTube comments most sensible people liked it though

Excellent essay + video! Wrote about some very similar observations over the weekend: Increasingly looking forward to a world where fewer developers feel like their abilities are starting to atrophy because of coding agents.

You rage against people saying you're all ragebait, and then post crap like this.

The challenge isn't the tech. It's the real-world impact.

since when tf do we measure the quality of software or an LLM by how many LOC you spit out per day?? people with zero real tech background (or the CEOs you’re talking about) only chase the stuff that gives instant dopamine raw LOC counts, pretty UIs, whatever looks impressive on a screenshot. even worse is when teams just ship whatever the model barfs out without reviewing a single thing. they’ve basically made their users the QA department now lol

Love this. Watching while Codex runs

as someone who have focused on building, this hits home. i've been struggling to get customers or clients to work with, and my social skills kinda suck right now. been chasing this "slop machine" since the Cursor starter days, and haven't stopped since. i have no fucking idea on what to focus on now.

CoDiNg WiTh AgEnTs Is A tRaP, aNd We AlL fElL fOr It! I aM tHeO. Amazing stuff - Much Thought.....Provoked. 👍👍 Two Thumbs!

a lot of people replaced learning with delegation way too early

Oh no, I can do test spikes with AI in 1 hour that would have taken me a week to code by hand, I'm so trapped, someone please help

Theo I’ve watched EVERY single one of your videos since you were blonde with stash and collabing with prime a bunch, you’ve hugely influenced my approach to building, but for the love of god ease up on the self glazing.

Bedrock accepted on T3 Code? That’s huge - if enterprise bedrock accepted please lmk

If you can use it with Codex, Cursor, Claude Code or OpenCode, you can use it in T3 Code :)

I am programming with ai pretty slow to be honest. Maybe slower than before. Most of the time goes into rubber ducking therapy sessions. Then I let it implement in phases and review each one before continuing. Coding speed is not my primary concern. The ability to think through a problem with a sparring partner is. That has definitely improved the architecture of my code by a lot. But in implementation I have to intenionally slow down. Let it only implement one phase at a time. That allows me to catch holes in the plan before they compound out of control.

I didn’t. I keep it slim and trim and it just works. The models are smart and all of the extras just burns tokens and context window

I think it depends on whether you genuinely enjoy what you are working on. Many people dislike the projects they are assigned to, so they simply paste the task into an AI agent. The smarter ones go a step further: they build a script that reads the task and delegates the work to AI agents for them.

@mattpocockuk wondering if you want to chime in here. For example, what do you think about SDD?

Coding with agents being a trap is exactly why our seed repriced. One partner saw the trap, called it retention, and asked if bugs can get board seats.

I did sounds the alarm to a lot of dark factory proponents that were causing more harm than good

What got me is realizing how much "agent productivity" was actually me reading and patching its output. Felt fast in the moment, but I lost the muscle memory for the problem. Started writing tricky parts by hand again and my mental model is sharper.

Thinking of agentic workflows as a casino only works if you own the casino

bro you need to chill out with I'm better than yall stuff it stopped me from watching the video to the end

Can confirm (I wrote about this >1 year ago). @Tobi, any update on this?