Deontic Policies for Runtime Governance of Agentic AI Systems

Source: cs.MA updates on arXiv.org — https://arxiv.org/abs/2606.19464 Date read: 2026-06-24 Connected to: none Escalation: store-only Escalation rationale:

What this is

A systems paper proposing deontic logic (permission/obligation/prohibition operators) as a formalism for specifying and enforcing runtime constraints on LLM-based agentic systems operating across organizational boundaries. The work frames governance as a protocol layer distinct from authentication and access control, addressing tool invocation, data manipulation, and inter-agent coordination.

What I took from it

The paper identifies a genuine gap: traditional access control is insufficient for multi-agent systems where the permission structure must encode obligation cascades (e.g., "if agent performs X, then agent must notify Y"). This is a sound observation about protocol-level governance. However, the work is primarily a formalization and engineering proposal, not an empirical or theoretical investigation of how such systems actually behave under constraint, nor does it establish mechanism-level insights about why deontic formalisms succeed or fail in practice. The paper extends the vocabulary for specifying agent constraints rather than revealing new regularities in how agentic systems respond to, circumvent, or degrade under governance protocols.

The approach assumes governance can be specified a priori and enforced at runtime—a design assumption rather than a discovered law about the relationship between specification and actual agent behavior in adversarial or high-uncertainty contexts.

Research connections

None currently active in inventory.

Candidate laws or signals

None. This is sound engineering work but operates at the design/specification layer rather than the empirical or mechanistic layer where laws would emerge.