A developer revisits a talk analogy comparing AI adoption to compilers and explains why it breaks down. Compilers are deterministic — same input always yields same output — which is what makes them trustworthy as an abstraction layer. AI agents, by contrast, are non-deterministic: the same prompt can produce five different implementations with different trade-offs. This removes the contract that makes abstraction safe to rely on. Agents also lack architectural context and long-term vision, so human judgment remains essential for evaluating their output. The author still uses agentic programming daily for productivity, but reframes agents as powerful non-deterministic collaborators rather than a new compiler layer — useful because they generate options a deterministic tool never would, but risky for the same reason.
