AI system principles

Grounding, scoped tools, confirmation and evaluation.

Every AI system U Framework builds follows these principles. They come from shipping assistants that work with real data — see Alfred.

1. Ground it

An assistant is only as reliable as the data it reads. Sources, freshness and access rules are designed explicitly. When a lookup fails, the system says so; it never substitutes a plausible guess.

2. Scope the tools

Tools are narrow and typed. A model receives only the tools it needs for the task, with inputs validated on the server.

TypeScript

const tools = [
  {
    name: "get_quote",
    description: "Latest quote for a symbol from the configured data provider.",
    input_schema: { type: "object", properties: { symbol: { type: "string" } }, required: ["symbol"] },
  },
];

3. Confirm consequential actions

Anything that changes state on a person's behalf — a trade, a message, a record — is returned as a proposal and executed only after explicit confirmation.

4. Show the work

Interfaces surface which sources and tools were used to produce an answer, so people can judge it.

5. Evaluate continuously

Behaviour is measured on task-specific evaluation sets before and after every change to prompts, models or tools. Regressions block releases.

6. Represent state honestly

Visual states — thinking, analyzing, responding — map only to real system behaviour. An interface never displays activity that is not happening.

2. Scope the tools

Tools are narrow and typed. A model receives only the tools it needs for the task, with inputs validated on the server.

TypeScript

const tools = [
  {
    name: "get_quote",
    description: "Latest quote for a symbol from the configured data provider.",
    input_schema: { type: "object", properties: { symbol: { type: "string" } }, required: ["symbol"] },
  },
];