Engineering
05
Overview
Where it’s going. The long-term plan is for people to create their own agents in Norma: one that logs in and reads transactions, one that moves money to savings on payday, or one that runs a routine they taught it by chatting. Eventually the infrastructure has to support many agents, built by many people, doing different jobs.
One reliable agent first. Before any of that, one agent has to do what it’s supposed to, the same way every time. If one agent can’t be trusted with money, more of them only multiplies the risk. Language models don’t give the same output every time, so the reliability has to come from the system around the model, its harness: what the model is allowed to do, what code checks, and what happens when something doesn’t add up.
Choosing the model. I plan to use one of the latest Claude models. So far, it’s been the most reliable model I’ve used in real work, but it’s also expensive. That cost means I need to test it thoroughly and confirm that people actually want Norma before it runs for real users.
Who does what. The harness starts by splitting Norma’s work into four parts, each handled by the right kind of tool. The more a job’s mistakes would cost, the less freedom it gets.
- Claude reads what people ask and writes the replies.
- Jev, a classification model, sorts transactions into categories and spots recurring bills.
- Plain code does all the math.
- Only code, plus the person’s approval, can move money within an account or between their accounts.