A Policy Algebra for Trust-Preserving Agentic AI Execution

Large language model-based agentic frameworks primarily optimize capability, such as reasoning, tool calling, and goal completion. However, enterprise execution requires reliable capability, ensuring that an agent does not achieve goals through unauthorized data access, widened authority, unapproved side effects, unrecoverable budget consumption, or incomplete evidence. This paper addresses how an enterprise platform can preserve useful agent capability while ensuring that every executed action remains authorized, non-amplifying under delegation, economically recoverable, and auditable.

The authors introduce a policy algebra that defines the reliability envelope within which agent capability may be exercised. Security profiles and runtime obligations compose through joins, intersections, budget narrowing, approval inheritance, and evidence accumulation. This composition is trust-preserving and forms the least restrictive state satisfying all governing inputs. Furthermore, the framework propagates restrictions across multi-agent calls and introduces cost-aware artifact materialization, which redirects open-ended execution toward a recoverable outcome as budget exposure grows.

The evaluation frames agent performance around a reliability-capability trade-off rather than a standard capability benchmark. The policy-algebra runtime intervenes on 94.8% of policy-violating events while maintaining an 86.9% task-completion rate. Additionally, the system eliminates observed profile-monotonicity and zero-artifact-exhaustion violations, while increasing audit completeness to 98.6%. These results demonstrate that formal correctness conditions and executable decision semantics can successfully govern agentic execution.