Agentic Inequality

Matthew Sharp, Omer Bilgin, Iason Gabriel and Lewis Hammond ask how autonomous delegates could redistribute power and opportunity. Their conceptual framework separates agent availability, quality and quantity. The concern is not only whether an assistant follows its owner: agents representing different people may compete, and unequal delegating capacity can change which people achieve their goals.^[1]

Access, capability and scale

In the October 2025 v2, availability is the foundational divide between being able to use a functional agent and being unable to do so. Access to a foundation model is insufficient when useful delegation also requires scaffolding, integrations, infrastructure and expertise. The authors include voluntary non-use alongside exclusion through economic or technical barriers, without treating their reasons as equivalent (§2.1).^[1]

Quality includes core intelligence, speed and throughput, reliability, tool use and behavioral dispositions. These properties need not move together: a strategically aggressive negotiator may be useful in bargaining but unsuitable for customer service. Quantity concerns how many agents an actor can deploy and coordinate, allowing parallel tasks beyond one delegate’s scale. The dimensions can compound; the benefit also depends on the user’s hardware, skills and other resources (§§2.2–2.4).^[1]

Why autonomous delegation changes the question

The paper distinguishes augmenting a person’s work from delegating an entire workflow. It reviews competing accounts of labor substitution and productivity gains, while warning that evidence from assistive tools may not transfer to autonomous agents. Similarly, firms might use proprietary data and many agents to reinforce a competitive advantage, yet lower-cost agent platforms could help smaller firms undertake work previously beyond their resources (§3.1). These are possible mechanisms synthesized from other research, not new causal estimates produced by this paper.^[1]

Its citizen-state example shows the distributional stakes. Premium agents could help affluent users navigate healthcare or legal bureaucracies, especially if services become optimized for agent interaction. Public-good agents could instead help people complete forms and pursue entitlements they would otherwise miss. In politics, scalable delegates could amplify wealthy actors’ influence or help citizens organize. The outcome depends on access and institutions, rather than on autonomy alone (§3.2).^[1]

Infrastructure determines usable power

Proprietary and open-weight models create different access barriers. A hosted service can hide infrastructure and maintenance costs from the user while creating dependence on a platform. Open weights permit direct adaptation but leave deployment expertise and computing costs to whoever uses them. The authors therefore reject the assumption that either release model is inherently more equitable (§4.1.2).^[1]

Tool access creates additional gatekeepers. Cloud providers, financial institutions and software platforms can determine which agents may use APIs, data and services. A capable model can consequently remain an ineffective delegate if excluded from the systems needed to act. The authors propose interoperability and open standards as possible responses, without demonstrating that these measures eliminate the underlying inequalities (§4.1.3).^[1]

Governing a distribution, not just a model

The agenda first asks how to measure availability, quality and quantity, then which disparities are unjust. It proposes participatory approaches to that normative question, infrastructure research, public-service agents and regulation of competitive agent interactions. “Universal basic agency” is a proposed baseline of autonomous assistance, with public provision and subsidies left as alternatives to investigate (§5.2).^[1]

Interventions have their own trade-offs. A capability ceiling might restrict useful innovation; compliance costs might drive providers from smaller markets; a universal public service built around one supplier might entrench vendor lock-in. The paper does not validate these policy effects or predict that inequality inevitably rises. Its contribution is to locate choices that disappear if equal access to the same model is treated as equal power (§5.1).^[1]

Agentic Inequality explains the broader idea. Social Alignment and Society-in-the-Loop: Programming the Algorithmic Social Contract ask how affected interests enter governance; this paper adds the capacity of those people to act. Aligned with Whom? Direct and Social Goals for AI Systems separates serving a principal from accounting for effects on others.

Versions and chronology

The first arXiv submission was October 19, 2025; this account uses the October 22 v2 revision, whose manuscript is dated October 21. Oxford Martin’s April 2026 report contains the April 24 v3 manuscript and substantively revises the framework and supporting literature. That later report remains a distinct source, rather than a mirror silently substituted for v2.^[1][2]