Sondera Compiles Natural-Language Rules into Provable Control Over AI Agent Actions
Sondera's research on autoformalization has gained recognition with its acceptance at the Agents in the Wild workshop at ICML 2026 and the LLM-Solve workshop at FLoC 2026. The company will also showcase a related tool demonstration at Black Hat Arsenal. The research focuses on transforming natural-language policies into formally verified rules for AI agents.
The paper titled "Autoformalization of Agent Instructions into Policy-as-Code," authored by Adam Mondl, Matthew Maisel, and John Brock, was recognized at ICML 2026, a prominent machine learning conference, as well as at the FLoC 2026 workshop dedicated to applying large language models to formal reasoning. The accompanying tool, "GolemHalt: A Deterministic Reference Monitor for AI Coding Agents," is set to be demonstrated at Black Hat Arsenal.
Utilizing peer-reviewed research with MedAgentBench, an independent benchmark for clinical AI agents published in NEJM AI, Sondera's pipeline has been shown to autoformalize a greater proportion of a complex 88-rule policy than previous hand-coded efforts. Specifically, the pipeline achieved a success rate of blocking all 99 adversarial unsafe write attempts.
Autoformalization plays a vital role in enforcing agent policies because actual policies are often expressed in natural language within documents like HIPAA manuals and standard operating procedures. Traditionally, converting these policies into enforceable rules has been a manual task, which is not sustainable given the rapid deployment of AI agents.
Sondera's pipeline automates this process by converting natural language directly into formally verified Cedar policy-as-code. A theorem prover reviews each rule, while adversarial simulations stress-test the rules before deployment. This neurosymbolic approach utilizes neural classifiers and symbolic rules to ensure that decisions made by AI agents are deterministic and tailored to the agent's overall trajectory.
Josh Devon, co-founder and CEO of Sondera, emphasized the need for robust policy enforcement, stating that current agent mishaps often stem from legitimate user actions rather than malicious intrusions. Sondera's policy and agent control plane is currently in private beta, and organizations interested in participating can find more information on the company's website.
The full research paper is accessible at arxiv.org/abs/2606.26649.