What an AI agent in production actually needs

The gap between demo and production

In a demo an agent has to be right once. In production it has to be right a thousand times in a row, and when it is wrong the mistake has to be traceable. That difference is not a model difference — it is infrastructure.

Tools, memory, permissions

An agent with no access to real tools only talks. Connections to internal APIs belong behind least-privilege scopes with full event logging, so every action stays auditable.

Evaluation before release

Every version needs a test suite: real questions, expected answers, and a metric that catches a quality regression before your users do.