It is not hard to make an agent.
You take a model, you give it tools, you write a loop, you post a screenshot. The internet will clap. I have clapped.
It is hard to make an agent that is still correct after the tenth tool call, after the retrieved document is stale, after the user asked for something the ontology does not contain.
What I want instead
I want:
- Schemas at the boundaries. Free text is how systems invent facts.
- Retrieval that can be inspected. If I cannot see the passage, I do not want the claim.
- Memory that is designed, not accidental chat history.
- Evaluation that is not "the answer looked nice."
- A failure mode that is quieter than hallucination and louder than a spinner.
The macro-regime work I did with LangChain, Ollama, and Weaviate was mostly an excuse to practice those constraints. The model was local. The ontology was small. The interesting part was refusing to let the generator write a market essay until the regime had been classified.
The business version
Companies do not buy agents. They buy a decision, a workflow, and a way to know when it stopped working. If your architecture cannot say those three things, you have a demo with an invoice.
I am still thinking about this. That is not a hedge. It is the state of the field.