@adlrocha - Fat agents vs. Narrow agents

Over the past year, I’ve built my own PR reviewer, an implementation agent, an email summariser, my own agent harness, my financial agent, and a good chunk of Baselight AI. In essence, I’ve been just building the exact same things everyone was building (except maybe Baselight) as a way to experiment with this new technology (and paradigm) that we like to call LLMs. Like everyone else, I’ve been learning as I go, because there are not yet any best practices or “the right way of doing it” (and if someone tells you otherwise they either lie or they are trying to sell you something).And I would expect that we all experienced the same thing, the first few versions of these agents we built by vomiting a big prompt (system or not) to the agent hoping for them to do what we said. This seemed to work… sometimes. Then different kinds of failures started to surface: an edge case here, a regression after a model update, a convention broken there, a run where the agent does the complete opposite…