Law School for Robots: Building AI Governance Stacks

Read full text Summarise with VIRAT
Authors: Daniel Rosehill, Gemini 3.1 (Flash), Chatterbox TTS
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Episode summary: As AI agents transition from simple chatbots to autonomous fiduciaries capable of moving capital and signing contracts, the industry is facing a critical challenge: how do we ensure these systems act within safe boundaries? This episode explores the shift from basic prompt engineering to "policy engineering" and the emergence of the Governance Stack. We dive into the March 2026 NIST guidelines on AI agent risk management and discuss why traditional system prompts are no longer enough to prevent catastrophic financial or legal errors. By implementing hierarchical document structures—comprising Constitutions, Bylaws, and Operating Guidelines—developers can create a more robust framework for machine reasoning. We also examine the technical architecture required to enforce these rules, including Retrieval-Augmented Generation (RAG) for policy fetching and the rise of "Auditor Agents" that serve as a digital check-and-balance system. Whether you are building autonomous trading bots or automated procurement systems, understanding how to encode human judgment into machine-verifiable constraints is the next great frontier in AI development. Show Notes The evolution of artificial intelligence has reached a critical tipping point. We are moving rapidly from the era of chatbots, where the primary risk is a "bad answer," to the era of autonomous agents, where the risk is a "bad action." When an agent has the authority to liquidate a portfolio, negotiate a contract, or manage a supply chain, a simple system prompt is no longer a sufficient safeguard. The industry is now pivoting toward a more robust framework known as the Governance Stack. ### From Prompts to Policy Engineering Traditional prompt engineering often treats all instructions with equal weight. In a single system prompt, a stylistic preference like "avoid using emojis" might compete for the model's attention with a critical financial constraint like "never exceed a $10,000 limit." Because of how LLM attention mechanisms work, these crucial hard stops can sometimes be deprioritized or "lost in the middle" of a long instruction set. To solve this, developers are adopting a hierarchical approach to governance. This "Governance Stack" mirrors legal and corporate structures, divided into three distinct layers: 1. **The Constitution:** High-level core values and the primary mission. 2. **The Bylaws:** Non-negotiable, binary rules and hard constraints. 3. **Operating Guidelines:** Tactical preferences, style guides, and day-to-day procedures. ### Architectural Enforcement Structuring the documents is only half the battle; the system must also be architected to respect them. Rather than stuffing every rule into a single context window, modern frameworks use Retrieval-Augmented Generation (RAG) to pull in only the relevant policies for a specific task. This keeps the agent's focus sharp and prevents context degradation. Furthermore, a mandatory "reasoning loop" acts as a gatekeeper. Before an agent executes an action, it must generate a reasoning block that explicitly maps its proposed move against the active policy. If the agent cannot justify the action within the established bylaws, the system triggers an automatic halt. This transforms the agent into a self-auditing entity that must "think" before it acts. ### The Role of the Auditor Agent One of the most promising developments in AI safety is the "Supervisor" or "Auditor" architecture. This involves a separation of powers where a Primary Agent performs the work, while a second, more constrained Auditor Agent reviews the output against the policy stack. This mimics the relationship between a CEO and a General Counsel. While the Primary Agent is focused on achieving the mission, the Auditor Agent is focused solely on compliance. This creates a digital check-and-balance system that can flag subjective issues, such as an agent becoming too aggressive in negotiations or drifting away from the intended corporate tone. ### Risk-Based Oversight Not every task requires the same level of scrutiny. Following the latest NIST guidelines, developers are moving toward risk-based governance. Low-stakes tasks might only require periodic sampling by an auditor, while high-stakes financial or legal actions require a "triple-check" and human-in-the-loop intervention. By defining the boundaries rather than the exact path, we can create AI agents that are flexible enough to navigate the real world but structured enough to remain under our control. Listen online: https://myweirdprompts.com/episode/governance-stack-autonomous-agents

Fields of Study & Subjects

Ethics and Social Impacts of AI Artificial Intelligence in Law Digitalization Law and Regulation

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
22 March 2026
Language
EN