Cognitive Offloading II

My feeling is that agents are mature enough to solve well-defined problems. Even if a problem is complex and hard, it’s still achievable either now or in the near future. Thus I’ve decided to lean into delegating to agents. As far as I’m concerned, agents are good at writing course-level projects. So I’m going to conduct a small experiment: expedite my understanding of computer systems, which is only missing one last piece, and push myself to learn more about machine learning at the system level, such as modeling and serving.

It looks like the hard part of implementation is offloaded to LLMs — but that’s only half the story. Agents can write any well-known data structure. The question of whether I should write algorithms and protocols by hand is gone. Modern system engineering treats LLMs as assistants to engineers, and staring at the screen for hours to identify an error log has become obsolete. It’s not that this skill is non-essential — it’s just replaceable. What gets freed up is time for high-level design and driving projects forward — and that’s the part that’s actually hard.

Building a good project no longer hinges on implementation skill. Good taste, programming habits, and review standards matter even more. Current LLMs are still not good enough for certain hard tasks where human judgment plays a central role. But building a good feedback system to drive agents forward is good practice. Agents should be able to learn from prior experience. However, letting them write their own memory is not enough nowadays — for either humans or agents. That’s where the project-doc skill (from Jinyan) comes in.

I agree with Brooker that it’s more important to establish the constants first — not letting agents do all the work, but laying the foundation on which to build the system components. The human decides the direction, not just the implementation details.

Even so, working with agents — even for just a couple of hours — is exhausting. It feels like competing with a powerful AI, and the question is: how do you find your own value in the collaboration? The answer is for you to find out. My plan is to get my hands dirty on real projects and have fun.