Autonomy is earned, not switched on: progressive validation before unsupervised agents

Edward C. Cheng et al. · 1 authors · Stanford University and Inquiryon, Inc. · 2026 · arXiv:2610.30445

Autonomy is earned, not switched on: progressive validation before unsupervised agents

Opening

The default move with a new agent is to grant broad permissions up front and add oversight after something breaks. This paper orders it the other way. Autonomy is something an agent earns through validation, and the structure that earns it has three parts: transparency, accountability, trustworthiness.

What the paper does

The authors present a Three-Pillar model and framework for developing safe and trustworthy AI agents, and note that existing human-in-the-loop and responsible AI approaches remain fragmented across system design, governance and operational deployment. The framework defines an evolutionary path in which increasing agent autonomy is introduced through progressive validation, structured human oversight, and auditable operational controls, then describes how the model supports lifecycle monitoring, decision traceability, escalation management, and human-agent collaboration in enterprise environments.

Existing Human-in-the-Loop (HITL) and responsible AI approaches provide important foundations, but many remain fragmented across system design, governance, and operational deployment.

Findings

  1. The three pillars are transparency, accountability, and trustworthiness — the framework develops safe and trustworthy agents by building those three together rather than adding oversight as an afterthought.

This paper presents a Three-Pillar (3P) model and framework for developing safe and trustworthy AI agents through transparency, accountability, and trustworthiness.

  1. Autonomy moves only through progressive validation — more autonomy is introduced through progressive validation, structured human oversight, and auditable operational controls.

The framework defines an evolutionary path in which increasing agent autonomy is introduced through progressive validation, structured human oversight, and auditable operational controls.

  1. The model covers the operational surface, not just principles — agent lifecycle monitoring, decision traceability, escalation management, and human-agent collaboration within enterprise agentic environments.

We further describe how the model supports agent lifecycle monitoring, decision traceability, escalation management, and human-agent collaboration within enterprise agentic environments.

Caveat

This is a framework, not a result: the abstract reports no deployment, no experiment, and no measured outcome, and offers no comparison against the fragmented human-in-the-loop approaches it describes. It also does not state how much oversight is enough at each autonomy level, so the promotion thresholds remain the operator’s to set.

What this means for 5ac

We already have the audit half: kanban events record every request, comment and completion, which is the decision traceability the framework asks for. What we lack is the written promotion rule. Write one: a new profile or card type starts fully reviewed, every result goes through a review handoff, and it earns auto-complete only after a stretch of clean passes. Demote back to review when a gate fails, and escalate after repeated failures instead of letting the same card cycle. That is progressive validation applied to bot ops, and it needs no new tooling.

Source