Governance frameworks for real-world deployment of agentic AI

Develop governance frameworks for real-world deployment of large language model-based agent systems, explicitly addressing long-horizon behaviors, planning-time failures, and multi-agent dynamics that are underexplored in existing benchmarks.

Background

The paper argues that embedding LLM agents into long, irreversible FAIR data management workflows introduces fail-open failure modes that require architectural governance. In surveying related work on agent frameworks and memory systems, the authors highlight that least-privilege access, immutable audit logs, and deterministic validation are often missing and that governance remains insufficiently addressed.

They explicitly cite a recent survey that frames the lack of governance frameworks for real-world deployment as a central open problem, especially given the limitations of current benchmarks which emphasize short-horizon behaviors over planning-time failures and multi-agent dynamics.

References

A recent survey identifies governance frameworks for real-world deployment as a central open problem, noting that existing benchmarks primarily focus on short-horizon behaviors, leaving planning-time failures and multi-agent dynamics underexplored [Wei et al. 2026].

— Exploring Robust Multi-Agent Workflows for Environmental Data Management  (2604.01647 - Guan et al., 2 Apr 2026) in Section 5 (Related Work)

Future work should evaluate the AI population workflow in domain-specific deployments, testing whether and how its population-level representations support institutional procedures across deployments and over time, regulatory requirements, and legal triggers.

— Convex AI Compositionality and the Governance of AI System Populations  (2609.24784 - Ferrario, 21 Sep 2026) in Section 6, “Final Recommendations and Conclusions”