Indian Researchers Propose Allocating AI Agent Authority Like a Budget

Who decides that a particular AI agent can keep running, and how much resource it is entitled to? In August 2026, three researchers from India who specialize in game theory and mechanism design answered that question in the paper "Resourced Authority: A Mechanism-Design Model for Participatory Governance of Deployed AI Agents" (arXiv 2608.06353). The authors are Praful Chandra, Sujit Gujar and Ganesh Ghalme. The MTS blog on Habr published a write-up of the paper on October 12.

The question left over after the risk debate

A great deal has been written about the risks of uncontrolled agents. These texts usually offer no practical fork in the road: the concern is justified, but who makes the decision here and now? Chandra, Gujar and Ghalme do not propose adding yet another list of prohibitions. They treat an agent's authority as a resource that someone must allocate to it. Hence the name: resourced authority, meaning authority backed by a resource.

The point is that an agent without a budget does nothing. Whoever controls the budget also controls the agent. If the budget decision is made not by a single developer but by the participants in the process, the result is participatory governance: collective management of systems that are already running.

An old idea in new packaging

The paper is full of heavy terminology such as coupling map and breadth-weighted effective supports. Judging by the write-up, the idea behind them is simple. It was devised more than a century ago to distribute money to public projects. The prototype of the debate over it is traced back to Athens, roughly two and a half thousand years earlier.

From the name "breadth-weighted effective supports," one can guess that the model counts not only the strength of support but also its breadth: how many different participants stand behind a decision. This is our interpretation. The formal definition and the proofs appear only in the paper itself, and the introductory part of the write-up does not reveal them.

What is known about the paper

• Title: Resourced Authority: A Mechanism-Design Model for Participatory Governance of Deployed AI Agents.
• Authors: Praful Chandra, Sujit Gujar, Ganesh Ghalme.
• Publication: arXiv 2608.06353, August 2026.
• Field: game theory and mechanism design.
• Subject: agents that are already deployed, not hypothetical superintelligent systems.
• Key concepts: coupling map, breadth-weighted effective supports.
• Source of the idea: mechanisms for funding public goods.

How engineers and managers should read this

For now this is an academic model, with no deployments and no empirical measurements, at least in the part of the material that is available. It is too early to count on a finished product.

Still, the paper is useful as a framework for thinking. If your company runs agents with access to money, code or infrastructure, it is worth asking yourself three questions:

• Does every agent have an explicit resource limit: tokens, API calls, write permissions, money?
• Who extends or cuts that limit, and by what procedure?
• Does an extension depend on the opinion of one person or of several stakeholders?

If the answer to the third question is "one person," mechanism design describes what could replace that. To judge whether the model works in practice, you will have to read the original: Habr points to the full text only behind its "Read more" link.