I tested the 'deterministic agent loop' claims with four experiments. They all failed — including my own fix.

A certain genre of "production-grade AI agent" article has been making the rounds. You know the shape: it argues that ReAct loops break in production, so you have to stack deterministic constraints on top of the LLM's uncertainty — a pre-AL gate, an LLM-as-Judge at temperature 0, a phase gate, a decision state machine. The one I have in mind claims 7000+ lines of production Rust.

The direction is right. Agent loops do need engineering guardrails; you can't let the LLM declare victory on its own. Pulling "self-contained agents" out of academic fantasy and toward engineering reality is a valuable move.

The problem is the repeated use of words like deterministic, objective fact, code vetoes the LLM to manufacture confidence. Do those claims actually hold up?

I didn't argue. I ran four experiments. Conclusion: each of the three core mechanisms it uses to establish "determinism" is only formally deterministic — all of them fail at the semantic layer. And the "upgrade" I prepared to fix them failed too.

Here's the data.