A good craftsman never blames his tools

There has been a tweet that has been in my thoughts a lot in the last.6 months – regarding AI agents. In the end, it boils down to “Agents are the Hello World of the AI Engineer.”

At first, I wasn’t sure about this sentiment. General agents may be some ways off, narrow agents can be hard to build correctly.

But there is a powerful thing that resonates with me here – that one of the most potent use cases that has panned out so far from the AI boom and Gen AI is the plethora of wonderful tooling that is available.

My own small example here was working on a small toy case with the Angreal repository. I built this last year as a way to pipe inputs from GPT-4 calls into my IDE, this was before Cursor.sh later became more prominent in my day to day workflow.

Agents seem like an even more powerful way to get leverage as an Engineer, and they also combine a very thin slice of the entire full stack of the AI Engineering world. RAG + Evals + Tool Use can give you a very useful narrow AI, and the power in much of Gen AI is not in true generality (Not yet anyway), but in dealing well in narrow domains that we deal with constantly.

The other exciting thing about Agents and other AI Dev Tools is that they themselves can be used to speed up the inner loop of development on the same project where they are being developed. I can’t think of anything that is greater catnip to developers (at least it is for me when working on these things.) than using a dev tool to speed up development of itself. It’s recursion, compounding gains and bootstrapping all rolled up into one.

While this mindset can always be dangerous from a build vs buy vs select scenario (it is easy to sink time more into tooling than doing things that generate value) – I think this is a facet of our industry that is exciting to me. If the right tool doesn’t exist in the world for what you want to do, you can build it yourself. Even if you don’t end up the top starred project on Github, you’ll learn a lot about the domain and likely become a better Engineer and improve your own workflow in the process.