Legora, a legal AI company, cannot see any trace of what its agent does for a lawyer, because the underlying data belongs to the law firms' clients. That constraint sits at the centre of a Latent Space podcast interview with Jacob Lauritzen, Legora's chief technology officer. The conversation addresses a hard engineering question: how do you build, debug and test an AI agent when you have no direct view of its work? For a company selling software into legal practices, where client confidentiality is a professional obligation, the usual approach of inspecting logs and traces is not available. The discussion covers how the team approaches development under that restriction. Lauritzen's role as CTO of one of the faster-growing companies in legal technology makes the topic relevant to other vendors facing similar limits. For Australian businesses and professionals, this changes little directly unless your firm already uses AI agents on client matters. If it does, ask your vendors how they test and monitor agents when they cannot access your data, and where that data is stored and processed. Client information is often personal information under the Privacy Act 1988 and the Australian Privacy Principles, and it may also be covered by professional confidentiality rules. Vendor testing practices are a reasonable part of any procurement review.