notes on ai agents
I had a tough time trying to figure out why someone would want to use AI Agents. At the end of the day, they are just LLM wrappers, right? Research labs are doing more important work: making datasets and training models. I was wrong!
AI agents are great, and when crafted with care, they can achieve amazing feats. They are able to do tasks more efficiently, save you costs and time (if you ignore the research time and cost, of course). The problem with the current ecosystem is probably that we are not trying to solve enough hard problems. That’s why building AI Agents seems futile. That’s why we hate prompting and are afraid of writing simple instructions because we fail to understand that hard problems can only be solved through the simplest approach.
We are trying to use MCP servers to buy a charger from Amazon, use 50 other MCP servers, and populate them with AI Agents to do web searches for us, write presentations, and make lightweight docs you can show your professors to get an A grade. But that’s not what AI Agents are meant for! These are valid use cases, but if we are always solving these non-trivial problems, we will never be able to appreciate AI Agents. More likely we are to get dumb ourselves.
I believe that even at their current capacity, AI Agents are great at solving hard problems. I’m not talking about cancer at this moment, or world hunger per se, but more complex problems than writing assignments, a problem that has real economic value. Be it military, finance, governance, or healthcare. Most problems in these areas are hard problems, and people would rather prefer easy alternatives with NLP and BERT-style classifications to navigate the problem space than actually try to build AI Agents. Why?
Because building real AI Agents is hard. That’s why you should build your own AI Agents because they are hard, and we are meant to solve hard problems. It might take a week, a fortnight, or a month, and there is a lot of uncertainty involved because you are trying to converge into a region of outputs where your agent should function. That’s okay. It takes time. But choose hard problems! Please don’t make a web-search AI Agent that can do homework for you. Instead, try to make an AI Agent that can help you figure out the right amount of theory and the right amount of questions that can make you do your homework on your own!
Try to solve hard problems, I guess?