Avatar Intelligence: Amplifying Human Agency Through Persistent Personal Agents

Avatar Intelligence. Amplifying Human Agency Through Persistent Personal Agents. By Zheng Wu and ten others, from Shanghai Jiao Tong University, the National University of Singapore, and one other institution. Published on the second of November.

Personal AI agents can increasingly handle long-horizon tasks across applications. The catch is on the human side. Users often have to put in substantial effort to specify goals, resolve ambiguity, and verify outcomes. An agent that can do more, but still needs to be told everything and checked on everything, has not saved its user much. The paper makes an argument about what progress should mean. The authors say it should be measured by how well agents amplify human agency: letting people pursue their goals with less routine supervision, while keeping control over the decisions that matter.

The paradigm they advocate is called Avatar Intelligence. These are persistent personal agents that represent a user's goals, values and constraints, through decision delegation sustained across contexts and over time. The paper lays out a maturity roadmap that runs from human-directed execution to autonomous operation within domains the user has delegated.

To evaluate where an agent stands, it introduces the Agentic Oversight Index, or AOI. It is a framework for weighing the value an agent delivers against two costs: the agent's own operational cost, and the burden of oversight placed on the human. An agent that delivers a great deal but demands constant supervision would score differently from one that delivers a little and asks almost nothing.

Three core research challenges are named. The first is identity-continuous personalization, carrying a user's preferences from one task to the next. The second is user-situated evolution, pursued through what the authors call Rehearsal Engineering. The third is safety, through mechanisms that pair technical control with accountability and recourse. What links them is persistence: an agent that remains the same representative of the same person, across tasks and over time.

The measure being proposed is not what an agent can do, but how much of a person's routine burden it lifts without taking away their say. Together these directions form a research agenda for reliable decision delegation to personal agents.

Avatar Intelligence. Amplifying Human Agency Through Persistent Personal Agents. By Zheng Wu and ten others, from Shanghai Jiao Tong University, the National University of Singapore, and one other institution. Published on the second of November. This is the long version: the whole paper rather than the abstract.

Hand a research task to a capable agent today and it will retrieve and summarize documents for you. You, meanwhile, are still resolving the ambiguous objective, correcting its interpretations, and judging whether the result is good enough. Execution effort goes down. The burden of specification and verification stays where it was. This position paper's answer is to change what counts as progress. The measure it proposes is how far an agent amplifies human agency: how many of your goals you can pursue without supervising every intermediate decision, while you keep control of the ones that matter. The authors, from Shanghai Jiao Tong University, the National University of Singapore and Shanghai University, call the paradigm Avatar Intelligence.

An avatar is a persistent personal agent that represents a user's goals, values and constraints across tasks and over time. The paper's own example makes the difference concrete. A work trip collides with a parent and teacher meeting the user promised to attend weeks earlier. A task by task assistant clears the calendar. An avatar first looks for a later departure, and if neither obligation can move, brings the conflict back to the user.

To place today's systems, the paper lays out a roadmap in four levels. At level zero the agent automates fragments and the user stays embedded in everything. At level one it runs longer workflows, but the user still steers key decisions. At level two an avatar core, meaning a persistent identity, memory and a decision policy, takes over routine decisions, and humans step in at a few calibrated checkpoints. At level three the avatar runs end to end within a delegated domain, and humans mostly audit afterwards. The level is per domain, so one avatar can sit at different levels for different parts of your life, and at every level the user can revise goals and revoke delegation.

The architecture is pictured as a user-centric operating system whose kernel is a continuously updated model of the user. Around it runs a loop of perception, planning, decision and execution, grounded in persistent memory. On top sit clones: work, assistant, social and custom, each a projection of the same identity into a different context.

The most concrete contribution is a score: the Agentic Oversight Index. Take delivered value minus the agent's own operating cost, such as compute and tool use. Divide by a task-specific scale and pass it through a hyperbolic tangent, so normalized net value lands strictly between minus one and one. Human oversight is counted separately, as a burden measured against a human doing the task alone: zero means no oversight, one means as much effort as doing it yourself. The index is normalized value minus a weight times that burden. The appendix builds the burden from intervention frequency, supervision time, the complexity of decisions demanded, and audit load, and defines agent cost to exclude it, so human effort is never counted twice.

The path from level two to level three, the paper argues, runs through three open problems. The first is identity-continuous personalization: deciding as the user would, not just in the user's preferred style. Stored memory is called brittle for agentic work, because coverage is incomplete, preferences depend on context, and people are not always consistent. The appendix separates how a user thinks from what a user knows, and asks for an identity that stays consistent across tasks without drifting.

The second is Rehearsal Engineering, a fourth layer beside prompt, context and harness engineering, which all act at inference time. Rehearsal works across episodes: collect the user's own traces, stage representative scenarios in simulation or sandboxes, evaluate candidate actions under explicit oversight policies, then update memory, skill selection and decision policy.

The third is safety in two layers. Technical safety covers corrigibility, least privilege, rollback and defence against prompt injection. Agent Insurance then allocates the residual financial risk through escrowed fees, collateral and underwriting, with claims settled on auditable execution traces.

There are no experiments. The paper builds no avatar, computes the index for no real system, and reports no results. What stands in for evidence is a survey of current products. Grok Bot keeps named bots and their context across tasks. Muse pursues ongoing goals through a persistent browser and asks permission before sending messages or making purchases. Cue imagines agents with their own email, phone and wallet. OpenAI's dots carries work between conversations. The paper places all of them at the move from human in the loop to human on the loop: its own roadmap applied to the market. The four application scenarios, personal, enterprise, financial and multi-avatar, are illustrations, not tests.

There is no limitations section. The risks the paper names live in its engineering comparison table, which lists drift, privacy and overfitting for Rehearsal Engineering, and in the safety section's admission that insurance offers recourse, not a guarantee of correct execution.

What the evidence cannot support is any claim about the index in practice. The appendix calls its functional forms illustrative, the weights are left to the evaluator, and no example score is worked through. Two avatars could rank differently under two reasonable choices of weight, and the paper offers no guidance on choosing.

The idea to carry away is the change of yardstick. An agent that finishes more tasks while still needing you to specify, check and rescue it has not necessarily given you anything. Score value net of the agent's cost, subtract your oversight as an explicit penalty, and let a bad agent sink below the cost of doing the work yourself. Progress, the authors conclude, should be shown through the decisions users can reliably delegate.