I love my job. That is exactly why this bothers me.
This is not one of those “software engineering is dead” posts, and I am not preparing to move to a farm because Claude wrote a working API endpoint. I have been interested in software since high school. Building things, breaking them, staying awake because there was one last problem I wanted to solve. That part has always been real for me.
But over the last year, something changed.
I get more done now than I ever did before. The strange part is that I enjoy less.
We became messengers for our agents.
I am no longer sure where the actual communication happens.
We say AI agents work for us. In practice, engineers increasingly act as messengers between them. One agent writes the implementation, another reviews it, and then a third checks security. You read their output, move comments between them, clarify requirements, reject unnecessary abstractions and ask for another iteration.
At some point, you are not really writing software. You are managing a conversation between machines.
The work still requires actual judgment. That part has not gone away. You need to understand the architecture, recognize a bad assumption, catch security problems and know when a technically correct implementation is completely wrong for the product.
But the actual experience feels different. You are not always creating anymore. Sometimes you are forwarding messages.
Engineers love getting lazy.
Or at least I do.
Engineering is largely the process of removing unnecessary work. We automate things because repeating the same operation feels stupid. Then we automate the automation because even triggering it manually starts to feel stupid.
First it was Copilot. Great. Less typing, fewer stupid repetitions, no real downside. You still understood the code, controlled the direction and stayed inside the flow.
Then agents started resolving issues from their own interfaces. That was also exciting, but copying code and context into a chat window became annoying. So we dragged them into the IDE.
Then came the next question. Why do we even care about the IDE?
Let the coding agents run directly through tmux. Give them the task, let them inspect the repository, implement the change and run the tests. The agent does the work. You read the result, point out what is wrong and make it try again. Still reasonable.
Then came another update. Copy your soul into a supervisor agent.
Give it your repository rules, your architecture preferences, the way you review code and all the small decisions you normally make without thinking. Let that agent coordinate several other agents. It assigns tasks, reviews implementations, requests corrections and reports back to you.
You review its review. Then you leave a few comments. Then the agents update the code, review it again and push it.
This is impressive. It is also fucking weird. What comes next?
Voice control, probably. I complain into a microphone while walking around the house and an agent converts my frustration into five pull requests. After that, maybe the system stops asking entirely. It learns my preferences well enough to recognize what I would reject before I see it.
At that point, what exactly am I doing? Waiting for agents to become slightly less annoying?
The dopamine disappeared.
The recent AI shift did not make me dislike software. It lowered the dopamine peaks I used to get from it.
Solving a difficult problem used to have a rhythm. You investigate. You get stuck. You understand one more part of the system. You try something. It fails. You try again. Eventually, the whole thing clicks.
The satisfaction came from getting there, not just from having the result.
Now, I can describe the problem to an agent, let it inspect the codebase, review the plan, correct two assumptions and receive a working implementation with tests before lunch.
Obviously, this is better for the business. I am not stupid enough to argue otherwise. It is also much less satisfying.
The problem disappeared before I had time to properly struggle with it. I know how ridiculous that sounds. Engineers spent decades building tools to avoid unnecessary struggle, and now I am complaining because they work too well.
But the struggle was not always wasted time. Some of it was the actual experience of engineering.
The AI slop problem.
Then there is the documentation.
Every company now has documents that nobody wrote and nobody wants to read. Ten pages explaining something a human could explain properly in two. Huge security reports where every minor concern is expanded into a dramatic finding. Task descriptions packed with objectives, expected outcomes, implementation notes and other filler that says almost nothing.
I sometimes read an AI-generated document and wonder whether anyone read it before sharing it.
I do not care that AI wrote it. I care that nobody cleaned it up.
The frustrating part is when somebody gives an agent ten seconds of effort and pushes the reading and cleanup onto everyone else. The work did not disappear. It was simply transferred to the reader.
At some point, passing untouched AI output to someone else starts to feel disrespectful. It no longer feels like communication. It feels like someone let their agent do the talking and forwarded the result without checking whether it sounded human.
We have made generating text almost free. Reading it is still expensive. That imbalance is turning a lot of work communication into garbage.
I am not competing with the agents.
I cannot compete with them on output, and luckily I do not need to.
I can design an architecture, divide the implementation into reasonable pieces, send several agents into the codebase, review their decisions, catch bad assumptions and get the whole feature implemented and tested in a day.
If I wrote every part myself, it could easily take a week.
That is real leverage.
It also means the valuable part of my work moved up a layer. I write less code myself and spend more time deciding what should get built in the first place, and how the pieces fit together.
That sounds great on paper. But there is a loss hidden inside that transition.
I still know how to build the thing. I am just not sure I enjoy the new way of building it.
Vibe coding is fine. Until it matters.
I do not have anything against vibe coders.
I like that people can now build things they could not have built two years ago.
More people can turn an idea into something real, and that is genuinely a good thing.
I also use the same approach for small games, experiments and side projects. Describe the idea. Generate a rough version. Play with it. Throw it away when it stops being fun.
But a lot of the vibe-coded applications I have seen are vulnerable, badly structured and not thought through beyond the happy path.
I would not keep critical personal information in most of them. I would not trust them with authentication, billing, private messages, financial data or anything else that would create a real problem when the generated abstraction fails in some unexpected corner.
The issue is not that the person did not manually type the code. The issue is that nobody understood the full system.
Agents are very good at producing something that looks finished. They are less reliable at recognizing when the entire direction is wrong, when a requirement is missing or when two individually reasonable decisions create a dangerous system together.
That instinct still matters, although I have no idea for how long.
Maybe I was worried about the wrong thing.
Since the AI paradigm shift began, I never seriously believed my job would simply disappear.
I saw AI as a tool for increasing efficiency, expanding markets and producing products much faster. I still believe that.
I was looking for the wrong threat. Maybe the real threat is not losing the job. Maybe it is losing the feeling that made the job worth doing...
The satisfaction of building something with other people. The long technical discussion that ends with a better design. The small moment when everyone sees the product working for the first time. The feeling that you created something together rather than supervised its generation.
That possibility bothers me more than unemployment.
So what now?
I don't have an answer here.
This is partly why I have a blog. If I cannot use it to dump an unfinished thought like this, there is not much point in having one.
These systems are becoming extremely powerful.
Perhaps the ideal future really is renting out 25 percent of my brain to orchestrate these systems unconsciously, leaving the rest of it for things I actually enjoy.
Let the agents inspect logs, fix regressions, write tests and argue with each other.
That would leave more space for tennis, friends, family, good coffee, actual travel, building something stupid just because it sounds fun, and hopefully learning enough to produce some music of my own, even if it is terrible.
Maybe I expected too much meaning from work in the first place. AI is just making that harder to ignore.
Still, I miss the excitement.
I miss feeling like I was building the thing, not managing the things building the thing. I do not know whether this is temporary or whether this is simply the job now. Right now, the work is faster and somehow less fun.