AI news ·
Researchers propose avatar intelligence to measure personal AI agents by human agency
A new paper proposes measuring AI agents by the oversight they demand, not tasks attempted. Its Agentic Oversight Index weighs value against run and supervision costs.
Personal AI agents can already book meetings, draft documents, and run multi-step research tasks. The harder problem is what happens after: users spend much of their time specifying goals, resolving ambiguity, and checking the agent's work. A new research paper argues that progress should be measured by how well these agents amplify human agency, not by how many tasks they can attempt.
The paper proposes Avatar Intelligence: persistent personal agents that represent a user's goals, values, and constraints, and carry decisions across contexts and over time. The authors call for sustained decision delegation, where an agent handles routine choices within agreed limits while the person keeps control of consequential ones.
A maturity roadmap, not a single leap
The paper lays out a maturity path from human-directed execution, where the user drives every step, to autonomous operation inside domains the user has explicitly delegated. The intermediate stages matter most for professional settings, where an agent might prepare a first draft or flag a conflict but leave the final call to a person.
To judge whether delegation pays off, the authors introduce the Agentic Oversight Index (AOI). It weighs the value an agent delivers against two costs: what it takes to run the agent, and how much supervision the user still has to provide. An agent that saves an hour of work but demands fifty minutes of checking scores poorly under that measure.
Three research challenges
The paper identifies three problems that have to be solved before persistent agents become dependable.
- Identity-continuous personalization. An agent needs to carry a user's preferences and context from one task to the next, rather than starting fresh each time.
- User-situated evolution through Rehearsal Engineering. Agents should adapt to the specific person and workplace they serve, improving through structured practice rather than generic retraining.
- Safety with accountability and recourse. Technical controls alone aren't enough. The authors argue for mechanisms that assign responsibility and give users a route to challenge or reverse an agent's decisions.
The framing matters for teams weighing where to spend effort. Most current tooling focuses on what an agent can do. The AOI shifts attention to the oversight load, which is often the hidden cost in any automation rollout. Organizations comparing tools may find that measure more useful than task counts. Courses on AI Agent Courses increasingly cover exactly this question of where human review belongs in an agent workflow.
What delegation means in practice
The paper's central claim is about control, not autonomy for its own sake. "Progress should be measured by how effectively agents amplify human agency: enabling people to pursue their goals with less routine supervision while retaining control over consequential decisions," the authors write.
That distinction separates Avatar Intelligence from the general-purpose assistant model. A persistent agent that knows a user's constraints can act without asking, but only inside boundaries the user has set. The research agenda is about making those boundaries reliable across tasks and over long periods.
Why this matters for [target audience]
If you work in a field where decisions carry consequences - a legal review, a patient triage note, a hiring shortlist, a grant application - the value of an AI agent depends less on what it can produce and more on how much of your judgment it leaves intact. When evaluating agent tools, ask two questions the AOI makes explicit: how much time does the agent actually save after you've checked its output, and can you trace and reverse a decision it made on your behalf? Teams building internal agent workflows can find related material in AI Productivity Courses.