Agentic coding is to human software development like a word processor is to handwriting. You can achieve similar output in half the time, but the gains of agentic development are exponential compared to the linear gains of typing versus handwriting. If you can see the path from A to B, the Agent can build you the car and pave the road for you to drive there in a few hours where a human would have to manually toil away for several weeks. For 90% of software development tasks, this is the right choice, but there remains that final 10% that divides the revenue-generating potential of software from the free (as in freedom) artistic pursuit of software.
Software is a means to an end. I use it to create products that are sold to customers, which generate revenue that pays my salary and I get to go home to a well-furnished apartment and eat three square meals a day. I don’t feel ownership over the code I write at work; it existed before me and dozens of hands have molded it over the course of its life. I follow the same language choices, linter rules, formatting rules, object-oriented paradigm, build tools, frameworks, and infrastructure that my company has imposed upon me. Sure, I could go make a big stink to write a service in Lisp, but that’s counterproductive to the goal of the job which is to ship the damn product and go home. To that end, agentic coding is perfect because it helps me get there that much faster.
Already, we can all run our own mini team of Agents in order to write lots and lots of code really quickly. With this level of orchestration across a whole team, we could be committing hundreds of pull requests a day. The bottleneck is no longer writing the code; instead, it is the human reviewer having to check the parts that the Agent cannot. The problem is, how does the human know that what the Agent is doing is correct?
Oftentimes, I will give a rubber stamp of approval to pull requests that I don’t have enough context on to give a good review since I don’t want to block my coworker. Plus, it’s not really up to the reviewer to catch bugs; the responsibility still falls on the original author. Now, with Agents writing the code, the original author is no longer the human and sometimes the human whose name is on the pull request doesn’t have full context on the problem they’re working on. This is particularly prevalent for junior engineers but can also affect anyone new to the codebase or the team.
In school, I had to put in the reps to get good grades. I had to do the homework problems, bang my head against the wall, ask questions, and work with my peers to develop my skills enough to pass the final exam. The same thing is true for learning anything; I have to spend enough time with the subject matter to develop my own understanding. Software development is no different. I started off web development with raw HTML before moving to jQuery and PHP, eventually landing my first job using React with Node.js. I had to put in the reps to learn what I know now. I’m far along enough in my career now that I have my coding fundamentals down but I am still developing my intuition on what makes a good product.
So much can be delegated to the Agent now. I don’t even have to read the tickets anymore. With enough MCP and API integrations, I can have the Agent create its own prompt by reading the Linear ticket and asking follow-up questions through Slack. Then it’ll code the solution, write tests to test the solution, then post the pull request without any intervention from me. If I’m not in the loop at all, how can I ever hope to spend enough time with the problem to develop intuition for what is correct?
Conversely, is asking an Agent questions about the code really that different from asking my coworker to explain something to me? That comes down to a matter of context. Does the limit of contextual knowledge start and stop in the codebase or is there something more that exists outside the grasp of the Agent that only my coworker can explain to me? Maybe there’s an edge case or specific design decision that is not clear to the Agent or, even worse, it hallucinates and gives me bogus information.
No one is forcing me to hand over my agency to the Agent. Unfortunately, there’s limited time in a day. Once my boss starts to expect the speed of output that only an Agent can deliver, it’s too little too late to reverse course. I’ve lost the opportunity to spend a lot of time with the code to learn things gradually. Now I must leverage the Agent even more to ask it questions about the codebase instead of doing the hard work of actually reading and tracing the code to draw my own inferences and understanding.
When I stopped speaking Chinese with my mother every day, my ability to recall started to degrade. I can feel myself struggling to find the words when I talk to her now. The same is true with software development. When I first started using Cursor, I felt like it could read my mind with its AI tab completion. But when I turned off my work computer for the weekend and came back to Neovim to edit this website, I realized that I was struggling to write basic CSS without Cursor’s AI tab completion.
I have to remind myself to treat Agents as tools to prevent myself from becoming overly reliant. It’s up to me to slow down so that I can build a strong foundational understanding of the problem space, the technology, and how it fits together in the product. None of that is guaranteed to happen for free anymore. It’s hard to keep up. At the end of a long day, when I still see a mountain of tasks to finish, when I have a deadline looming on the horizon, when I just want a bit of room to breathe; in those moments of weakness, it’s easy to let the Agent run amok. Maybe it’s correct 90% of the time, and maybe that’s good enough. What scares me, though, is that I am not certain it is correct. At least when I write the code, even if it’s buggy, I can be certain that, at the time, I believed it was correct. So while I may no longer be writing the code myself, I have to force myself to take the time to prove the Agent’s correctness.
There is a part of me that still yearns for the tactile feel of writing my own code with a keyboard. I love my side projects. I get to experiment with different tooling, different paradigms, and different languages. I love chasing down novel ideas and experimenting to learn or just for the joy of writing well-thought-out programs. There was a time before all the Agents that I would do this at work too, but that time has passed now.
Agentic coding is like the fashion industry. There’s Shein, just pumping out trends, as fast as fast fashion gets. Then there’s Evan Kinori, hand-crafting every piece, focusing on the materials, slow fashion at its finest.
I think software and clothes are just a means to an end, but I respect that for some people, the means is the end. - Nathan Chau
The slow fashion equivalent for the software world would be the open source folks. They build slowly, creating reliable systems that must work. They’re not beholden to the timeline of a company and the expected returns of its shareholders. While not completely shielded from Agents, and even finding benefit in using Agents, they can scope their policy to prevent the recklessness that AI can bring upon an organization.
Lately, product managers, designers, and even support teams have all started to contribute code. They’ve always hated the bottleneck of engineering time and now there’s a tool to circumvent that. The engineers don’t mind either since it lets them focus on more interesting tasks than having to adjust the padding by two pixels or change the color of the text. At the end of the day, the engineers still have to review the code and correct any AI slop that non-engineers create. As this paradigm proliferates, the pull requests from product managers and designers will start to get bigger and bigger until the engineers are all but janitors for the AI slop that the rest of the organization outputs.
Maybe someday, perhaps someday soon, the Agents will be good enough that none of this will be a problem anymore. The code will write itself and be perfect. The Agent will manage its own orchestration; we will simply assign it tickets and it will create the perfect system. Perhaps we will have to build many, many more data centers to see that day come and face the environmental consequences.
My agents drank more water than me today. - Michael Zhao
Then engineers will really be obsolete, or perhaps move into some other role that the Agent still can’t do, and keep moving roles as our numbers dwindle. Is it stupid to think that there is a unique human essence that cannot be replicated by machines when we are inherently creating things for the machine? Does that human essence even have to come from someone who understands computers? Does anyone actually even understand the inner workings of an LLM’s reasoning? Perhaps this is the great democratization of technology and there really is no artisanal slow fashion equivalent for software and all this pining for a simpler time will forever be forlorn. But I’m hopeful for the future. Prometheus gave man fire and Google gave man “Attention Is All You Need”. Fire can burn but at the same time, it gives us light and provides us warmth. Some will be burned but the march of progress continues.