When the Baseline Wins
A recent research paper built an AI-agent system to authorize access requests and compared it with a simple rule-based evaluator. The rule-based evaluator was more accurate and roughly 100,000 times faster.
What the researchers built
In Agentic-ZTA: A Multi-Agent Architecture for Autonomous Zero Trust Enforcement (arXiv:2610.05782, October 2026), Shovan Roy, Lopamudra Praharaj, Maanak Gupta and Bhavani Thuraisingham describe a system in which specialized AI agents assess each access request (identity, permissions, device integrity and policy) and a trust algorithm combines their judgments into an allow or deny decision. It follows the zero trust model in NIST SP 800-207.
What they found
Across 160 synthetic requests, half of which should have been denied, the agentic system reached 95% accuracy. It wrongly granted 3 of the 80 requests it should have refused, and wrongly blocked 5 of the 80 it should have allowed. A deterministic rule-based evaluator, run on the same requests, was correct every time.
Each agentic decision took between about 45 and 134 seconds. The rule-based evaluator decided in under a millisecond. The authors conclude: “The current prototype therefore does not support real-time enforcement.”
They also report that they did not measure repeated runs on identical inputs because model inference is not deterministic.
Why it matters
The study is small and synthetic, and the authors say so. But it separates two jobs that are often blurred together.
Interpreting a situation (gathering evidence, weighing context, spotting what’s unusual) is where AI can genuinely help. Deciding whether an action is authorized is different. That decision has to be correct every time, fast enough to keep up with the action, and consistent on Tuesday as on Monday. An auditor has to be able to re-run it.
In this experiment, the deterministic evaluator did all three. The AI did not, and the authors deserve credit for publishing the comparison that shows it.
The practical lesson for anyone deploying agents: let AI inform the decision, and keep the decision itself deterministic.
Triarcus Systems builds Cathedral, a deterministic governance architecture for autonomous systems. Cornerstone is its working implementation. info@triarcussystems.com