On AI Agents, the Freedom to Slow Down, and Accountability in an Age of Automated Knowledge
Recently, I’ve been experimenting with some of the newer AI Agent products, and I’ve found myself having mixed feelings about them.
On one hand, I’m genuinely impressed by what they can do. AI is no longer just a chatbot that answers questions. It can autonomously plan tasks, operate tools, interact with software, and complete work on my behalf in the background.
My life has become more convenient.
But on the other hand, I’ve noticed that my cognitive burden has actually increased.
I find myself worrying about what data an Agent has accessed, how much permission it has, whether my private information is secure, and how closely I need to supervise its actions.
And perhaps more importantly, I’ve started wondering: if everyone else is using these Agents and I’m not, am I falling behind?
Something that was supposed to simplify my life has introduced an entirely new set of things to worry about.
1. AI Automates the Workload, but Also the Joy
I’ve always believed that automation should help us get rid of the things we don’t want to do.
But many AI products seem to operate under a different assumption: if something can be automated, the entire process should be automated.
It reminds me of cooking.
I would love to have someone wash and chop the vegetables, prepare the ingredients, and clean the dishes afterward.
But I want to do the actual cooking myself.
I enjoy hearing food sizzling in the pan, watching the ingredients gradually transform, and deciding when to add seasoning or turn off the heat.
That process is part of the pleasure of cooking.
If a robot told me that cooking fumes were bad for my health and therefore decided to prepare the entire meal for me, I might end up with a faster, safer dinner.
But I would also lose the experience I wanted in the first place.
AI has automated the workload, but it has also automated the joy.
And this creates a strange problem.
Once the activities that brought me joy have been automated, I have to find new activities to enjoy.
Technology keeps saving me time, while I keep searching for new ways to fill that time.
I’m not against automation.
Sometimes, I genuinely enjoy letting an Agent complete an entire task for me and watching how it plans and executes the process. Experiencing the technology itself can be fun.
But I want to be able to choose.
Today, I might ask an Agent to order coffee for me. Tomorrow, I might want to walk outside, explore the neighborhood, research local cafés, and enjoy the process of choosing one myself.
The freedom to automate should also include the freedom not to automate.
2. AI Is Increasing Productivity, but Reducing Autonomy
This feeling becomes even stronger at work.
With the arrival of AI Coding Agents, companies have started raising their productivity expectations for engineers.
If AI can generate code faster, shouldn’t a task that previously took five days now take two?
If one engineer can accomplish more with AI, shouldn’t that engineer be expected to take on more work?
On the surface, companies are simply providing employees with more powerful tools.
But when higher productivity expectations follow, using AI gradually changes from a choice into an obligation.
Technically, I can choose not to use AI.
But I may no longer be able to meet the new performance expectations without it.
I am no longer choosing to use AI. I am being expected to use AI.
That creates a strange kind of pressure.
Technology was supposed to expand my choices, but now it seems to be narrowing them.
I no longer feel that I have the freedom to slow down.
I can’t simply decide that today I want to spend more time understanding a problem, studying an algorithm, or debugging a piece of code myself.
Because AI can already do those things for me.
But those are precisely the things that made me want to become an engineer.
I enjoy understanding why a system runs slowly. I enjoy profiling a workload from software all the way down to the hardware, identifying the actual performance bottleneck, and designing experiments to validate my hypothesis.
If all of those processes are automated and my role is reduced to approving the final result, I might accomplish more work without gaining more satisfaction from my career.
3. We Are Scaling Code Generation Faster Than Engineering Understanding
AI Coding Agents are also fundamentally changing software development workflows.
Traditionally, engineers developed their understanding while writing code.
Designing, implementing, testing, and debugging were interconnected parts of the same process.
Now, AI can generate enormous amounts of code in a very short time.
An engineer may not have enough time to understand everything it produces.
Under delivery pressure, we may end up skimming the implementation, deciding that it looks reasonable, and trying it out.If it works, great.
If it doesn’t, we ask AI to fix it.
Then we ask an AI Reviewing Agent to review the code generated by the AI Coding Agent.
The workflow becomes:AI generates code → AI reviews code → AI fixes code → Human approves.
The problem is not that AI is necessarily more error-prone than humans.
The problem is that the speed of code generation is beginning to exceed the speed at which humans can understand and verify the code being generated.
When organizations focus primarily on delivery speed, engineers may gradually shift from being creators to being supervisors.
But supervision does not necessarily imply understanding.
An engineer clicking “Approve” does not mean they understand the design decisions behind the implementation. Nor does it mean the code has been sufficiently verified.
This raises a question for me:Are we actually improving software engineering productivity, or are we simply increasing the speed at which code is produced?
4. But Humans Have Always Worked Through Trial and ErrorThen again, the more I think about it, the less convinced I am that AI Coding has introduced an entirely new problem.
Human engineers have always worked through trial and error.
When we encounter unfamiliar problems, we search Google, read Stack Overflow, look at implementations on GitHub, and copy, modify, run, and debug code.
The information we find is not necessarily correct.
Human knowledge does not appear out of nowhere.
We have always relied on existing information, generated new evidence through experiments, and continuously revised our understanding.
Scientific research works in much the same way.
We propose hypotheses, design experiments, collect data, analyze results, and formulate new hypotheses.
If AI can perform these steps faster and continuously produce useful new experimental data, then the knowledge-production process may not be fundamentally different from the one humans have always followed.
AI may not have created the verification problem. It may simply have amplified a problem that has always existed.
When the cost of generating knowledge falls dramatically while the cost of verifying knowledge does not fall at the same rate, the problem becomes increasingly visible.
5. Peer Review Has Never Been Complete Verification
This reminds me of my own experience reviewing papers for AI conferences.
As submission volumes increased, I received more and more review requests.
In reality, I simply did not have enough time or computing resources to independently reproduce every experiment in every paper.
Most of the time, I could only read the paper and evaluate whether the research question was meaningful, the methodology was reasonable, the experimental results supported the conclusions, and the overall argument was convincing.
If I found the topic interesting, the methodology plausible, and no obvious problems with the work, I might give the paper a positive review.
But that did not mean I had independently verified every claim.
I suspect many other researchers face similar constraints.
Peer review has never been equivalent to independent reproduction.
A paper being accepted does not mean its conclusions have been proven correct.
In a sense, academic research has always allowed researchers to publish new findings and let subsequent work test, extend, challenge, and refine them.
So if we insist that every piece of AI-generated code or every AI-generated research finding must undergo complete independent verification before entering the broader body of knowledge, we may be holding AI to a standard that human knowledge production itself has rarely been able to meet consistently.
The real question should not simply be whether AI makes mistakes.
It should be:How do we build a system that can continuously discover, correct, and learn from errors?
6. Reputation Is a Form of Long-Term Accountability
But none of this means researchers should care less about rigor.
Quite the opposite.
I believe researchers should hold themselves to a high standard when publishing their work.
Not because peer reviewers will necessarily catch every mistake.
Not because every irreproducible paper will be retracted.
But because academic research is a long-term process of building credibility.
Suppose I publish a paper on Efficient AI, claiming that my method significantly improves model inference speed.
Other researchers trust my findings, cite my work, and spend GPU resources and research time trying to reproduce my results.
If they repeatedly find that my method cannot be reproduced, and I cannot provide a reasonable explanation, they may become less willing to rely on my research in the future.
Even if the paper is never retracted, my academic credibility may gradually suffer.
Of course, a single failed reproduction attempt does not necessarily mean a study is flawed. Citation counts are not a direct measure of research quality, either.
But over the long term, whether a researcher’s work is reliable and genuinely useful to others can influence whether people continue to trust and build upon it.Reputation is accumulated through repeated interactions, not granted by a single acceptance decision.
That is why rigor should not come solely from external review.
It should also come from a researcher’s own commitment to their long-term credibility.
I can use AI to help me generate hypotheses, write code, run experiments, and perhaps even carry out most of the research process.
But if the final work is published under my name, I should be able to explain my technical claims, understand the evidence supporting them, and honestly communicate their limitations.
That does not mean I have to perform every step personally.
It means I should take responsibility for the work I choose to put my name on.
7. The Freedom to Choose How We Participate
This brings me back to where I started.
I am not against AI Agents.
Nor do I believe humans must personally perform every task to prove their value.
If AI can write correct code faster, conduct experiments more efficiently, and discover new knowledge more reliably, those capabilities deserve to be used.
But I want technological progress to genuinely expand human freedom.
I want to use Agents when I care about efficiency and participate personally when I care about exploration.
I want companies to benefit from AI-driven productivity gains while still allowing engineers the space to understand, create, and learn.
And as AI becomes increasingly capable of generating knowledge, I hope we can build more effective verification systems rather than simply transferring ever-growing review responsibilities to already overwhelmed humans.
Most importantly, I want people to retain ownership of their choices and their work.
AI should expand our capabilities without taking away our autonomy.
Technology can perform work on our behalf.
But it should not decide for us which activities are worth participating in, which processes are worth enjoying, or how quickly we should live our lives.
Ultimately, the goal is not to turn life into an increasingly efficient production line.
It is to gain more capabilities while retaining the freedom to choose the life we want to live.
Siqi, Etc. 