TL;DR
This paper introduces Avatar Intelligence (AvI), a framework for persistent AI agents that act as digital representatives of a user’s identity, values, and goals. The authors propose moving beyond simple task completion toward "agency amplification," measured by the new Agentic Oversight Index (AOI), which evaluates an agent's value while explicitly penalizing the human effort required to supervise it.
What Problem Does This Paper Address?
While current AI agents are becoming adept at long-horizon tasks, they still suffer from a high "supervision tax." Users must spend significant time specifying goals, resolving ambiguities, and verifying outputs. The authors argue that current research focuses too much on raw execution quality and not enough on decision delegation.
The core challenge is creating agents that can operate autonomously across different contexts and over time without needing constant "hand-holding." To solve this, we need a way to quantify when an agent is actually saving a user effort versus just shifting the type of work the user has to do.
Methods
The paper outlines a conceptual and mathematical framework for Avatar Intelligence rather than a single new algorithm. Key components include:
- User-Centric Operating System: A proposed architecture where a shared "Avatar Core" (containing an identity model and long-term memory) supports various task-specific "Clones" (e.g., work, social, assistant).
- Maturity Roadmap (L0–L3): A four-stage progression from "Human-for-the-loop" (tools) to "Human-out-of-the-loop" (avatars), where the agent assumes end-to-end autonomy within specific delegated domains.
- Agentic Oversight Index (AOI): A novel evaluation metric defined as $AOI = U - \lambda P$. Here, $U$ is the normalized net value the agent provides, $P$ is the human oversight burden, and $\lambda$ is a weight representing the user's sensitivity to being interrupted.
- Rehearsal Engineering: A methodology for specializing agents by having them "rehearse" scenarios in simulated or sandboxed versions of a user's specific environment to learn their unique trade-offs and preferences.
Datasets
The paper is a position paper and research agenda; therefore, it does not introduce or use specific training datasets. Instead, it references the trajectory of foundation models like GPT-5.5 and Gemini-3 (projected into the year 2026) as the "backbone" upon which these avatar systems will be built.
Key Insights and Ideas
- Agency Amplification vs. Automation: The goal isn't just to automate tasks, but to allow humans to pursue higher-level goals with less routine supervision.
- Computable Identity: For an avatar to be effective, a user's cognition must become "computable." This involves Thought Computability (representing how a user reasons) and Intent Predictability (anticipating what a user wants without being told).
- The "Clone" Paradigm: Instead of one giant model doing everything, the authors suggest specialized clones that inherit a central identity. This allows for "Work Continuity" across different apps and timeframes.
- Agent Insurance: Acknowledging that technical safety isn't perfect, the authors propose "Layer II" safety: a financial infrastructure for agent insurance to compensate users for losses incurred during delegated tasks (like financial trades or code deployments).
Key Results
As a theoretical framework, the "results" are the formalizations provided in the appendices:
- Formalization of Personalization: The authors define the personalization objective as a predictability constraint, ensuring the avatar's behavior remains consistent with the user's identity across different tasks.
- AOI Instantiation: The paper provides a mathematical model for the oversight burden ($H$), aggregating factors like intervention frequency, supervision time, and cognitive complexity of decisions. It demonstrates that $AOI$ converges to net value as oversight drops to zero, but can become deeply negative if an agent is too "needy."
Limitations and Future Work
The authors acknowledge several significant hurdles:
- Identity Drift: As both the user and the avatar evolve, maintaining a "stable yet updateable" identity model is difficult.
- Privacy in Multi-Avatar Ecosystems: When avatars negotiate with each other (e.g., two avatars trying to schedule a meeting), they must reach agreements without leaking too much of their respective users' private preferences.
- Safety and Recourse: Moving to L3 autonomy (Human-out-of-the-loop) requires "institutional-grade assurance." The transition from execution-centric assistance to high-stakes decision delegation remains the most significant research gap.
- Overfitting in Rehearsal: There is a risk that agents might overfit to specific "rehearsed" scenarios and fail to generalize when the real-world environment changes slightly.
Summary updated