All education is about learning how to think. It is about learning to problem solved. Whether that was 30 years ago, now, or 30 years from now, I think it is the same.
Technology changes over time, but if you can learn how to solve problems, you are going to do great things.
This blog is my reflections from the past few months of working with coding agents, especially Claude Fable.
Coding-bound
Most systems PhD students I know were coding-bound. If we break down the time it takes to produce a paper, most of it isn’t spent on writing or plotting, though both are extremely important. It goes into coding, debugging, and profiling. I also think this is one reason a systems paper takes a year (before the AI era). Systems venues are full of some of the best coders in the world, and even reviewers weigh the engineering effort behind the work.
“Bound” isn’t a negative word here. I believe it’s how we train our skills: before you can (over)polish a round ball, you first have to make one which is roughly round. Beyond skills, there’s the joy of finding bottlenecks, solving problems, and helping real people and real projects. That’s one reason I enjoy research and am still doing it.
Not coding-bound anymore
I would not go lengthy here. Coding agents are simply much better than me in several ways. They write code faster than I do. They debug faster than I do. They profile impressively well. They also know far more public information than I do. I can’t match their speed on these tasks.
Will coding agents replace systems PhDs?
This is the elephant in the room. My current answer is no. Being coding-bound meant that implementation took most of my time; it didn’t mean implementation was my most important skill or contribution in a PhD.
I believe a good systems PhD researcher should be able to:
- Identify a problem worth solving: explain who needs a solution, why existing approaches fall short, and why now is the right time to work on it.
- Develop and evaluate a solution, even including solving problem from other domains.
- Communicate the work through writing, plots, and presentations, so others can understand it and build on it.
Coding agents can help with all three. But I still need to decide whether the problem matters, whether the evidence supports the proposed solution, and what others should take away from the work. Faster implementation gives me more room to make those decisions and explore more problems.
My real internship mentor at Microsoft, one of the best vector search researchers, told me to keep exploring and ideation. That advice feels even more relevant now that trying an idea takes less time. To me, that curiosity is central to research, especially applied research in computer science. A 3Blue1Brown video I recently watched about studying math also got me thinking about the role of coding. I see a parallel with calculation in mathematics: fluency matters. But doing the calculations faster doesn’t, by itself, tell you which questions are worth asking or solving.
Information-bound
My routine used to be: code until tired, rest, repeat. That’s no longer the case thanks to coding agents. Lately I’ve had more room to explore papers, systems, and possible research directions (due to my SoftwareCampus)—something I always wanted to do more of, but had little time for while I was coding-bound.
The consequence of all this exploring is that I’m now bound by information consumption. I feel limited in how much information I can absorb and understand deeply. Reading a paper or getting an explanation from an agent is only the beginning. I still need time to understand the details, connect what I’ve learned to my own work, and decide what is worth pursuing.
What comes next?
I’m a longtime reader of Viktor Leis, who once said that “for systems, research is development and development is research”. I still think this holds. It’s just that I’m no longer the one writing the code. But I am still developing.
Looking ahead, I don’t feel ideation-bound yet. A PhD doesn’t need (too) many ideas, just a single-digit number of good ones. The harder part for me now is learning enough to judge which ones deserve the time.