AI Writes More Code, but the Developer Is More Exhausted Than Ever

Last year, I joined a new project.

The team adopted a new way of collaborating: whenever possible, hand the work to AI. People would mainly be responsible for setting requirements, making judgments, and accepting results.

This was very different from how I used to work. Before, I wrote the core logic myself and let AI help fill in code and fix bugs. On this project, AI took over most of the implementation work, and I was responsible for the direction, quality, and final outcome.

After more than a month, I noticed something a bit unusual.

AI was clearly writing more code, and my daily output was higher than before, yet I felt more exhausted when I got off work. My brain felt stuffed with cotton. I had been busy all day, but looking back, I felt I hadn't accomplished much.

I talked to a few friends who also use AI heavily, and they had similar feelings. Later, I replayed the experience and found that the fatigue mostly came from the following things.

01 The scope of what I could do expanded, but my judgment didn't keep up

Before, I mainly wrote frontend code and only handled simple CRUD on the backend. With AI, I could produce something decent in a completely unfamiliar domain within a single day.

This experience makes it easy to overestimate yourself. AI temporarily lets you handle more work, but your judgment and experience don't grow at the same time. Code that runs doesn't mean you understand the risks, boundaries, and maintenance costs in that domain.

If you forget this, it's easy to misjudge timelines, take on tasks beyond your ability, or even make promises you can't keep.

AI can expand your range of action. Whether something should be done, how long it should take, and to what standard — those still require human judgment.

02 Waiting made me switch tasks constantly

When using AI, feedback comes quickly, but not so quickly that you can completely ignore the wait.

Type in a few prompts, watch it follow its own train of thought, and it's easy to get absorbed. Once a task starts running, I can't help but stare at the progress, feeling like the result is about to appear any second.

But those few minutes are awkward. Keep waiting, and it feels like a waste of time. Switch to something else, and just as you get into a flow, you have to come back to check the result. Switch a few times in a single morning, and while execution time might be saved, I could never focus on finishing one thing properly.

In the past, the fatigue from coding usually came from sustained thinking. Now there's a new kind of fatigue: the brain constantly starting, pausing, and then reconnecting to the previous task.

03 The work rhythm always gets stuck on "almost done"

Before, I would line up the day's tasks in the morning, push through them in order, and wrap up on time.

With AI, I often feel like I can finish a day's work in half a day, so I cram more tasks into the plan. In practice, a single feature might take half a day of debugging to meet expectations, and near the end of the day, there's still an agent running.

At that point, it's hard to just shut it down.

"The result should be out any moment, just wait a little longer."

After the result comes out, you still need to read the code, run tests, and handle new issues. By the time this round is done, dinner time has already passed.

AI shortened a lot of execution time, but it also blurred the boundary of when work ends. As long as there's one task still running, I feel like the day isn't wrapped up yet.

04 Reviewing multiple agents simultaneously is mentally draining

To improve efficiency, I would run several agents at the same time. They handle different tasks and progress at different speeds.

I need to check each result one by one, and every time I switch, I have to recall a whole set of context. Which files did this agent change, why was it designed this way, where is the test coverage, and where is the other task stuck — all of this has to be recalled quickly.

Reviewing also means making decisions continuously. Is the approach acceptable, does the risk need to be addressed now, should the issue be reworked or deferred to the next round. The decisions are small, but there are a lot of them.

Switching like this dozens of times a day — the fatigue isn't surprising.

05 Output increased, but the sense of completion weakened

This part is the hardest to notice.

Before, when I wrote a piece of core logic by hand, the journey from being stuck to getting it working was clear. Now, AI generates the code, and I'm responsible for describing, waiting, checking, and correcting. The project is indeed moving forward, but I find it hard to point to any part and say I completed it myself.

When the part you complete yourself shrinks, the sense of completion weakens along with it. After a busy day, having committed plenty of code, I still feel like I'm just continuously receiving results and processing results.

People need more than just output. Understanding a problem, making a choice, solving a difficult point with your own hands — these processes tell us where today's time was spent.

I started adjusting like this

During this period, I've been trying a few simple approaches.

Form your own judgment first, then let AI act.

Before fixing a bug, take a look at where the problem might be. Before writing a requirement, clarify the goal and acceptance criteria first. Having a basic route in your mind gives you a basis for reviewing AI's results and makes it harder to be led astray by it.

Set aside a period each day without AI.

This time can be used to read code, organize tasks, or fill in knowledge gaps. Being able to understand and advance the project without AI makes using it feel more grounded.

Batch AI tasks together.

Try to send out work that needs agent execution in batches, and check them all at a fixed time. This reduces the urge to stare at progress and keeps the time in between intact.

Only accept code you can read.

AI's implementation can be different each time, and sometimes it will tear everything down and start over. A project needs long-term maintenance, so I at least need to know where the key logic is, why it's written that way, and where to start looking when something goes wrong. Code you can't read, even if it runs for now, will sooner or later become a new burden.

I now care more about one thing. At the end of the day, have the tasks truly been closed out, and do I clearly know what judgments I made? If I can answer these two questions, then the time AI saved truly belongs to me.