Human-Centered Adaptive Autonomy for Safe and Accountable LLM Agents via Conformal Risk Control

Tool-using LLM agents increasingly act on behalf of people inside workflows where an unsafe or unauthorized action can create financial loss, expose sensitive information, or bind an organization to an unwanted commitment. The core safety question is how to let an agent act under bounded delegated authority while preserving human oversight, accountability, and auditability. We propose an adaptive autonomy governor that treats autonomy as a continuous, enforceable budget, including spend caps, tool-permission scope, and irreversibility limits, and that can be tuned to an operator-set tail-risk tolerance $\alpha$ for the event that a high-risk action proceeds without review. The governor combines multi-axis risk signals spanning epistemic, economic, legal/compliance, and organizational-trust considerations with conformal calibration to produce (i) autonomy budgets, (ii) role-based review routing decisions, and (iii) structured audit logs. A procurement-negotiation example shows how calibrated authority allocation can keep routine edits moving while routing consequential actions to accountable human reviewers.