Research seriesL2paper
Permission models authorize an agent to touch a resource; they say nothing about why it should. This series builds intent-based governance from the ground up: governing an autonomous agent by its declared intent rather than an enumerated list of permissions. It develops the intent-governance model and how it differs from access control, how to capture and verify intent with intent specifications and provenance while confronting the intent-action gap, how to compile intent into enforceable runtime guardrails and where those guardrails fail open, how to measure intent conformance — did the agent do what was asked, and only that — and how to govern at scale with intent ledgers, audit trails, and a mapping to the NIST AI RMF, the EU AI Act, and ISO/IEC 42001. Grounded in the NIST AI Risk Management Framework and Generative AI Profile, NIST SP 800-207 Zero Trust, the OWASP Agentic Security Initiative and LLM Top 10, MITRE ATLAS, ISO/IEC 42001, and the EU AI Act.
Permission models say what an agent may touch; they are silent on why. Intent-based governance authorizes by declared purpose — and asks of every action whether it serves that purpose.
Governing by intent only works if the intent is captured faithfully, trusted to be authentic, and honest about what it fails to cover.
A declared intent that only lives in a document governs nothing; it has to compile into runtime guardrails that intercept every action — and you have to know exactly where those guardrails fail open.
Conformance is two questions, not one — did the agent accomplish the declared goal, and did it do nothing beyond it — and measuring the second is where governance is hardest.
One agent's intent is a design detail; a fleet's intents are a governance system — and it only holds together when every declared intent and every decision becomes durable, queryable, auditable evidence.