Sovereign Agentic Loops: Decoupling AI Reasoning from Execution in Real-World Systems
Jun He, Deying Yu
Why It Matters
What makes this one worth your time
This work is crucial for AI engineers and researchers focused on deploying AI systems safely, as it addresses the risk of unsafe executions by ensuring that AI intents are validated before affecting real-world systems.
SAL architecture enhances AI safety by validating intents before execution in real-world systems.
Summary
The paper introduces Sovereign Agentic Loops (SAL), a control-plane architecture designed to decouple AI reasoning from execution in real-world systems. SAL ensures safety by validating AI-generated intents against system state and policy before execution, using an obfuscation membrane and an Evidence Chain for auditability. The architecture is demonstrated in a cloud infrastructure prototype, showing effectiveness in blocking unsafe intents and preventing unsafe executions.
Key contributions
- Introduction of the Sovereign Agentic Loops (SAL) architecture for decoupling AI reasoning from execution.
- Implementation of a control-plane architecture that validates AI intents against system state and policy.
- Demonstration of SAL's effectiveness in a cloud infrastructure prototype, blocking unsafe intents and preventing unsafe executions.
Notable insights
- The use of an obfuscation membrane to limit model access to sensitive state is a clever approach to enhance security.
- The cryptographically linked Evidence Chain provides a robust mechanism for auditability and deterministic replay.
Possible limitations
- Not stated in the abstract
Abstract
arXiv:2604.22136v1 Announce Type: cross Abstract: Large language model (LLM) agents increasingly issue API calls that mutate real systems, yet many current architectures pass stochastic model outputs directly to execution layers. We argue that this coupling creates a safety risk because model correctness, context awareness, and alignment cannot be assumed at execution time. We introduce Sovereign Agentic Loops (SAL), a control-plane architecture in which models emit structured intents with justifications, and the control plane validates those intents against true system state and policy before execution. SAL combines an obfuscation membrane, which limits model access to identity-sensitive state, with a cryptographically linked Evidence Chain for auditability and replay. We formalize SAL and show that, under the stated assumptions, it provides policy-bounded execution, identity isolation, and deterministic replay. In an OpenKedge prototype for cloud infrastructure, SAL blocks 93% of unsafe intents at the policy layer, rejects the remaining 7% via consistency checks, prevents unsafe executions in our benchmark, and adds 12.4 ms median latency.