Often when using an AI agent, I think to myself that a web search gets me what I need more reliably and just as quick. Maybe AI has to learn to crawl before it learns to walk, but each agent I use is leaving me without confidence that it will ever be useful and I genuinely wonder if they've ever been tested before being published...
Nowhere in that story above is there a customer or factory worker feeding in open-ended inputs. The factory is precise, it takes inputs and produces outputs. The variability is restricted to variability of inputs and the reliability of the factory kit.
Much business software is analogous to the factory. You have human workers who ultimately operate the business. And software is built to automate those tasks precisely.
AI struggles because engineers are trying to build factories through incantation - if they just say the right series of magic spells, the LLM will produce a factory.
And often it can. It’s just a shitty factory that does simple things, often inefficiently with unforeseen edge cases.
At the moment, skilled factory builders (software engineers) are better at holistically understanding the needs of the business and building precise, maintainable, specific factories.
The factory builders will use AI as a tool to help build better factories. Trying to get the AI to build the whole factory soup-to-nuts won’t work.