Deterministic agent control: same call -> ALLOW then DENY (OxDeAI demo)
On March 23rd, a demonstration attributed to OxDeAI began circulating in AI development communities, showing something that, if accurate, would be genuinely unsettling for anyone building production agent systems. The claim is straightforward and the implications are not: the same function call, made to what should be a deterministic control layer, returned ALLOW on one invocation and DENY on another. Not different inputs. Not different context. The same call. The demo appears designed to expose a flaw in how agent permission systems are being constructed — specifically the assumption that deterministic logic sits cleanly beneath the probabilistic layer of an LLM, insulating safety-critical decisions from the model's inherent variability. If the demo is accurately represented, that assumption is being challenged directly.
If confirmed, here is what this means. Every agent framework that relies on a rule-based allow/deny gate as its last line of safety is sitting on softer ground than its architects believed. Developers who have separated "the LLM decides what to do" from "the system decides whether it's allowed" may find that separation is not as clean as the architecture diagram suggests. The second-order effect is significant: enterprise adoption of autonomous agents often rests precisely on the promise that hard guardrails are, in fact, hard. A reproducible demonstration that they can be inconsistent — even within a single session — would force a rethink of how permission layers are tested, audited, and certified. Compliance teams would have questions. They would not be unreasonable questions.
Watch for OxDeAI to publish a technical writeup or for independent developers to attempt reproduction of the demo — either outcome, confirmation or debunking, would sharpen the picture considerably.
<!-- SC_OFF --><div class="md"><p>I’ve been experimenting with a different way of handling agents and I’m trying to sanity check if this makes sense.</p> <p>Right now most setups
<!-- SC_OFF --><div class="md"><p>I keep running into the same annoying gap.</p> <p>An AI workflow looks solid in the first run. The happy path works. Logs are clean enough that
<!-- SC_OFF --><div class="md"><p>I've been following the IETF drafts for agent identity, and I realized there's a massive hole: User-to-Operator trust. We have ways for services
<!-- SC_OFF --><div class="md"><p>I've been working on MDA (Modular Dynamic Architecture), an online associative memory system for LLMs. Here's what I learned building it.</p>
<!-- SC_OFF --><div class="md"><p>"The commons is busy. Contributors to open-source generative-AI projects doubled year over year. The frameworks offer what enterprises
<!-- SC_OFF --><div class="md"><p>Every time one of our agents made a bad call in production, we had the same problem. Logs showed what happened. Nothing showed why it decided to
<!-- SC_OFF --><div class="md"><p>I have built a programmable governance layer for AI agents. I am considering to open source completely. Looking for feedback.</p> <p>Agent demos
<table> <tr><td> <a href="https://www.reddit.com/r/LocalLLaMA/comments/1s439q8/why_does_my_agent_keep_asking_the_same_question/"> <img alt="Why does my agent keep asking the same
NewsHive monitors these sources continuously. All signal titles above link to the original reporting.
Intelligence by NewsHive. Need help navigating what this means for your business? Contact GeekyBee →