In software, the code documents the app. In AI, the traces do.

In traditional software, code serves as the documentation for an application's logic, but for AI agents, this paradigm shifts. The actual decision-making for agents occurs within the model at runtime, making the code mere scaffolding for orchestrating LLM calls. Consequently, the true source of truth for an agent's behavior is found in its traces, which document the sequence of steps, reasoning, and tool usage. This change fundamentally alters development practices, requiring debugging, testing, and optimization to focus on analyzing traces rather than just code. Building effective AI agents therefore necessitates robust observability and structured tracing to understand and improve their non-deterministic behavior.