Researchers propose Avatar Intelligence to make AI agents better at handling your daily tasks

What you could do with it

• Researchers could evaluate whether a personal AI agent truly helps users by weighing the value it delivers against the human effort required to supervise it, using the Agentic Oversight Index ¹.
• AI developers could guide the design of persistent personal agents by following a maturity roadmap that moves from human-directed tasks to autonomous operation in delegated domains ¹.
• Users could track whether their AI assistant saves more time than it demands by applying a metric that balances agent output against operational cost and oversight burden ¹.

Figure from the paper: A detailed diagram showing applications of Avatar Intelligence including Enterprise Avatar, Personal Avatar, Financial Avatar, and Multi-Avatar Interaction with various AI assistant icons and interfac

Figure: Wu et al., arXiv:2611.03092, CC BY 4.0

Words to know

• Persistent agent: An AI assistant that remembers your preferences and goals over time, rather than starting fresh each session.
• Rehearsal Engineering: A proposed method for helping agents evolve and adapt based on real user situations through practice.

Researchers including Zheng Wu, Xinbei Ma, Yao Yao, and others published the paper on November 2, 2026, arguing that the AI industry should shift how it measures progress ¹. Instead of focusing only on what agents can do in isolation, they say success should be judged by how effectively personal agents amplify human agency — enabling people to pursue goals with less routine supervision while keeping control over important decisions.

The proposal introduces a maturity roadmap that moves from human-directed execution to autonomous operation within delegated domains. This means agents would start by helping with small, supervised tasks and gradually take on more responsibility in areas users explicitly delegate to them ¹.

To track whether an agent is actually delivering value, the team created the Agentic Oversight Index (AOI), a framework that weighs the value an agent provides against two costs: the operational cost of running the agent and the burden placed on the human to oversee it ¹. The idea is simple — if an agent requires as much work as it saves, it is not helping.

The researchers identified three core challenges that must be solved for reliable personal agents. First is identity-continuous personalization, meaning the agent must carry your preferences and knowledge consistently across different tasks and over time. Second is user-situated evolution through Rehearsal Engineering, which would allow agents to learn and adapt from real user experiences rather than being statically programmed. Third is safety, combining technical controls with accountability so users have recourse when something goes wrong ¹.

The paper was uploaded to Hugging Face and has received 12 upvotes from users in the AI research community ¹. It is positioned as a research agenda rather than a working product, outlining directions for building agents that can reliably handle decision delegation across tasks and over time .

Key facts

• Avatar Intelligence is a research framework, not a product
• Proposed by researchers at Shanghai Jiao Tong University
• Measures success by how well agents amplify human agency
• Introduces the Agentic Oversight Index (AOI) to track value vs. oversight cost

Sources

1. Paper page - Avatar Intelligence: Amplifying Human Agency Through Persistent Personal Agents