Avatar Intelligence: Amplifying Human Agency Through Persistent Personal Agents

The paper proposes Avatar Intelligence: persistent personal agents that represent a user's goals, values, and constraints, with a user-centric architecture, an L0-L3 delegation roadmap, and the Agentic Oversight Index for evaluating delivered value against human oversight burden.

Foundation models and agent frameworks make persistent personal agents increasingly plausible, but the paper argues a central question remains unresolved: how much work users must still do to make delegation effective. Execution effort may decrease while specification, coordination, and verification burden remains substantial. The paper's position is that personal agents should amplify human agency by reducing the effort needed for reliable, user-aligned outcomes while preserving users' ability to set and revise goals, choose delegations, and redirect action.

The paper characterizes Avatar Intelligence through persistent user representation and sustained decision delegation. It describes a user-centric operating-system architecture with a persistent user model and memory, a perception-planning-decision-execution loop, task-specialized clones, and three system capabilities. It outlines an L0-L3 maturity roadmap for transferring operational and decision responsibilities within delegated domains. It introduces the Agentic Oversight Index, which combines normalized net value U with an explicit penalty for human oversight burden P weighted by lambda. It identifies three research directions: identity-continuous personalization, user-situated evolution through rehearsal, and safety mechanisms for controlled and accountable delegation.

• The maturity roadmap defines four levels, from human-for-the-loop L0 to human-out-of-the-loop L3 autonomy within delegated domains, with users retaining authority to revise goals and revoke delegation at every level. §2.2 Roadmap
• The Agentic Oversight Index is proposed to assess capability improvements jointly with oversight burden, using normalized net value U and a penalty for human oversight burden P. §2.3 Agentic Oversight Index (AOI)
• The paper identifies three core research directions: identity-continuous personalization, user-situated evolution through rehearsal, and safety mechanisms for controlled and accountable delegation. §1 Introduction
• The avatar architecture is organized around four tightly coupled components: perception, planning, decision, and execution, grounded by a persistent memory substrate. §2.1 Avatar as a User-Centric System
• The paper lists four application settings: personal life, organizational work, financial decisions, and coordination among multiple avatars. §3.2 Application Scenarios
• Current industry products such as Grok Bot, Muse, Cue, and OpenAI's dots carry work forward over time while returning consequential approvals to users, marking a transition toward human-on-the-loop oversight within delegated tasks. §3.1 Current Industry Products

• The AOI functional forms are explicitly illustrative; the paper states only that net value is normalized and human oversight is represented once as a monotonic penalty, leaving concrete validation open.
• The paper presents a position and agenda; it does not report experiments validating the L0-L3 roadmap, the AOI, or the proposed architecture.
• Rehearsal Engineering is described with risks of drift, privacy, and overfitting and requires user- and environment-specific traces, staged scenarios, evaluation, and updates under bounded autonomy.
• Memory-heavy personalization heuristics are described as brittle for agentic tasks because coverage is incomplete, preferences are context-dependent, and user behavior may be internally inconsistent.

• Avatar Intelligence transforms agents from passive instruments into autonomous digital representatives that embody user identity, preferences, and values and operate persistently across contexts and time. §2 Avatar Intelligence
• The kernel of the user-centric system is a faithful, continuously updated model of the user's cognitive identity, not a thin memory cache. §2.1 Avatar as a User-Centric System
• Task-specialized clones for work, assistant, social, and custom contexts inherit the same user model while interfacing with tools, applications, sandboxed environments, and external agents. §2.1 Avatar as a User-Centric System
• The L0-L3 roadmap stages a decoupling of human involvement from agentic capability, and the same avatar can operate at different levels across domains. §2.2 Roadmap
• The AOI evaluates agent maturity through Agentic Intelligence and Human Involvement, connecting delivered value with the human effort required to obtain it. §2.3 Agentic Oversight Index (AOI)
• Existing personalization largely optimizes behavioral outputs rather than the latent cognitive model, motivating identity-continuous user representations that are stable yet updateable. §4.1 Personalization: From Behavioral Alignment to Identity-Continuous Avatars
• Rehearsal Engineering aligns a standardized agent to the user's ecology by collecting traces and preferences, staging scenarios, evaluating actions with oversight policies, and updating memory, skill selection, and decision policy under bounded autonomy. §4.2 Evolving with Specialization
• Avatar safety combines technical safety, which enforces delegated authority and limits failures, with Agent Insurance, which allocates residual financial risk under contractual terms. §4.3 Safety