Intent-Governed Loops for Accountable Agentic AI

AAAI 2026 Workshop TrustAgent Submission64 Authors

Keywords: Agentic AI systems, AI accountability, Runtime admissibility, Trustworthy AI, Decision provenance, Symbolic verification, Semantic verification, AI safety, Human-in-the-loop

TL;DR: An intent-governed loop converts alignment into admissibility enforcement with dual-mode enforcer and a temporal governance graph.

Abstract: Agentic LLM-based systems are now operating with tool access and institutional authority in finance, clinical triage, municipal policy, and similar high-liability domains. We argue that verifiable accountability in such settings requires architectural support beyond alignment: every emitted action must be provably authorized by an explicit mandate, within declared scope and constraints, based on current evidence, and escalated to named authority when outside mandate. We propose the intent-governed loop as a conceptual architecture that provides structural mechanisms toward these accountability properties through runtime control. This position paper articulates the core components: an Intent object (human context, symbolic constraints, semantic guidance); a Planner that proposes actions with structured justification; a dual-mode Enforcer that deterministically checks symbolic constraints then semantically evaluates boundary cases; and a temporal governance graph that records provenance, constraint evaluation, temporal coherence, and escalation. We outline key loop-level invariants, identify the non-optional architectural principles any implementation must satisfy, and propose an evaluation agenda including synthetic benchmarks, metrics, and adversarial stress tests to guide future empirical validation. Our position is that such an architecture represents necessary structural support for accountable deployment in high-liability domains, and we identify key architectural questions requiring further analysis before implementation.

Submission Number: 64