SJTU, NUS Paper Pitches 'Avatar Intelligence' for Personal Agents

A new position paper from researchers at Shanghai Jiao Tong University, the National University of Singapore and Shanghai University argues that progress in personal AI agents should be scored not by task completion but by how much supervision the user still has to do. The paper, posted to Hugging Face on 2 November 2026, proposes a paradigm it calls Avatar Intelligence: "persistent personal agents that represent users' goals, values, and constraints through sustained decision delegation across contexts and time."

Its central instrument is the Agentic Oversight Index, defined as AOI = U − λP, where U is normalised net value and P is a human-oversight penalty aggregating intervention frequency, supervision time, decision complexity and audit load. The authors sketch a maturity roadmap from L0 ("Human-for-the-loop") through L3 ("Human-out-of-the-loop"), with an "avatar core" of persistent identity, memory and decision policy taking over at L2.

The paper then names three open problems. First, personalization has to move "from surface-level preference matching to learning an identity-continuous model" that generalises across tasks. Second, agents need what the authors call Rehearsal Engineering — a "user-situated specialization loop" that stages scenarios in the user's own tools and routines rather than public benchmarks. Third, safety is split into two layers: runtime technical controls, and "Agent Insurance," a contractual mechanism that would escrow service fees, require posted collateral, and compensate covered losses from "fund-moving, code changes, and account operations" against auditable execution traces.

The paper situates itself alongside shipped products including Meta's MUSE, OpenAI's dots, xAI's Grok Bot and Cue, and lands in a week of unusually heavy agent coverage — the 526th agents-tagged story we've tracked in the last ninety days.