MIT’s Ramesh Raskar envisions decentralized, verifiable AI agents for AGI

A different answer to the AGI question

The current debate over artificial general intelligence often begins with a familiar assumption: if models become large enough, trained on enough data and tuned with enough compute, they will eventually cross a threshold into broad, general capability. Ramesh Raskar of MIT is challenging that premise from a different angle. In a newly published Forbes interview and analysis, he argues that AGI, as commonly marketed, is too vague and too centralized a concept to describe the next phase of AI development [1].

Raskar’s core claim is not simply that AI progress is slowing or accelerating. It is that the “AGI” label hides the real engineering and governance problem. The world’s knowledge is fragmented, private, regulated, proprietary and context-dependent. No single model can absorb all of it, especially when crucial knowledge sits inside pharmaceutical labs, hospitals, governments, companies, households and personal histories [1]. If that is true, then the future of capable AI may depend less on one universal system and more on many specialized agents that can prove what they are, negotiate with one another and act under verifiable constraints.

That is why Raskar’s proposal matters. It converts AGI from a question about one model’s intelligence into a question about an ecosystem. Instead of asking when a single system will become generally intelligent, he asks what infrastructure would allow billions of specialized, personal and institutional agents to collaborate safely across networks [1].

From foundation models to agent infrastructure

Raskar uses a transportation analogy to separate two layers that are often blurred together. Foundation models are like engines; agents are like cars; and the internet of agents still needs roads, traffic rules, testing facilities, registries and safety institutions [1]. The analogy is useful because it shows why model capability alone does not solve the deployment problem.

A model can generate text, write code, reason through a plan or call tools. But an agent that acts on someone’s behalf raises different questions. Who authorized it? What domain does it claim to serve? Is it certified for medical, financial or legal advice? Can its credentials be revoked? Can another agent check its claims in real time? Can users distinguish a trustworthy personal agent from an impersonator? Raskar’s view is that these questions belong to “agent safety,” not just “model safety” [1].

That distinction could reshape how companies build AI products. If the model is only the engine, then the most important business layer may move toward agent harnesses, verification services, registries, reputation systems and domain-specific safety scaffolding. Raskar specifically suggests that third-party players, not necessarily the companies that train foundation models, may provide the frameworks that industries such as pharma use to deploy agents [1].

This is a direct challenge to the vertically integrated AGI story. In that story, the company with the best model becomes the dominant AI provider. In Raskar’s story, the value shifts toward a decentralized market of interoperable agents, where the winning systems are not simply the smartest but the most reliable, auditable and useful in context.

NANDA and the “internet of AI agents”

The Forbes piece identifies Raskar’s work with Network AI Agents in Decentralized Architecture, or NANDA, a protocol intended to let agents work together across the internet [1]. Recent summaries of the same story describe the proposal as a vision of decentralized, verifiable and personalized AI agents collaborating across networks [2]. Another aggregation of the article uses similar language, emphasizing that Raskar is questioning current AGI assumptions rather than merely forecasting a bigger model race [3].

The key word is “network.” A single agent can be helpful, but a network of agents raises the possibility of coordination: a personal learning agent could interact with a school agent; a health agent could exchange limited, permissioned information with a clinic; a research agent could locate a certified domain expert agent; a business agent could negotiate with a supplier agent. But this only works if agents can discover one another, identify one another and verify claims.

Raskar’s list of implementation challenges is therefore infrastructure-heavy. He names four major needs: a registry for agents, attestation, verification and testing [1]. Those four terms define the practical agenda behind the vision.

A registry answers the discovery problem: where does an agent live, and how can another agent find it? Attestation answers the identity and credential problem: who says this agent is what it claims to be? Verification answers the trust problem: can its claims, permissions and behavior be checked continuously? Testing answers the safety problem: how should agents be evaluated before they operate in high-stakes environments? The Forbes Liberia version of the story highlights the same four issues, presenting them as the central obstacles to building the web of tomorrow [4].

Continuous verification, not one-time certification

One of Raskar’s most striking ideas is that AI agents may need continuous certification rather than a one-time approval. In the Forbes account, he gives examples of agents claiming to advise diabetics or provide financial guidance; such claims, he argues, should be attested and repeatedly verified rather than accepted at face value [1].

That proposal reflects the unusual nature of agentic AI. A conventional software product can be tested, released and patched. An AI agent may adapt its behavior, use external tools, interact with other agents and operate in changing contexts. If it acts in health, finance or education, a static certificate may not be enough. Its current model, permissions, data access, policy compliance and recent behavior may all matter.

This is where Raskar’s thinking departs from the familiar “trust us” model of AI deployment. The trust relationship is not supposed to rely solely on a brand, a model card or a user’s intuition. It would be mediated by technical and institutional systems that can check an agent’s status repeatedly. In practical terms, that could mean agent credentials, real-time compliance checks, reputation signals, audit logs, revocation mechanisms and controlled environments for testing risky behavior.

The idea also reframes regulation. Regulators often ask whether an AI system is safe enough to release. Raskar’s framework asks whether an agent remains safe enough to keep acting. That difference is crucial if agents become persistent digital representatives for individuals, firms and public agencies.

Edge AI and the personal-agent thesis

Raskar’s vision is not only decentralized at the network level. It is also decentralized at the hardware and data level. In the Forbes interview, he discusses India as a major example, saying the government is preparing to provide AI agents to citizens and that many agents can run locally on phones, laptops or PCs rather than depending entirely on massive data centers [1].

The claim is significant because it pushes against the idea that advanced AI must always mean larger cloud infrastructure. Raskar’s argument is that many useful agents can operate “on the edge,” close to the user and the user’s data [1]. This matters for privacy, cost, latency and personalization. A personal agent that learns from a user’s documents, habits, preferences, language and local context may not need to send every interaction to a remote data center.

That does not mean cloud AI disappears. Raskar acknowledges that cloud replicas or backups may play a role [1]. But his emphasis is on an architecture where personal agents are not merely thin clients attached to a central intelligence. They are crafted, local, user-specific systems that can carry some of the user’s expertise and preferences into the wider agent network.

This is also why the story connects to the idea of “8 billion agents.” Raskar imagines a future in which personal agents are highly individualized, reflecting people’s accumulated work, talent, context and priorities [1]. The point is not that every person gets the same chatbot. The point is that every person could have a distinct digital representative.

The bazaar era of AI

The most important implication of Raskar’s argument is economic. If AI becomes a network of agents, then the market does not revolve only around model subscriptions. It also requires identity rails, payment rails and a “justice system” for agents that misbehave [1].

That phrase may sound speculative, but it points to a real governance gap. If agents negotiate, buy, sell, recommend, schedule, diagnose, tutor or transact, then disputes will follow. An agent may misrepresent credentials, exceed authority, cause loss, leak data or manipulate another agent. The infrastructure must therefore include ways to assign responsibility, reverse transactions, suspend agents, resolve conflicts and protect users.

Raskar calls this the “bazaar era” of AI, contrasting it with a future in which users simply buy intelligence from a large centralized company [1]. A bazaar requires standards. It also requires pluralism. Agents from different builders, countries, sectors and communities must be able to interact without all belonging to one platform.

That is why decentralization is not just a philosophical preference in his argument. It is a functional requirement. A world of personalized agents cannot work if every agent must be owned, hosted and approved by one company. But decentralization without verification would be chaotic. Raskar’s formula is therefore not decentralization alone; it is decentralized coordination with registries, attestations, verification and testing.

Why this challenges the AGI race

The conventional AGI race rewards scale: more parameters, more compute, more data, more capital. Raskar’s framework rewards infrastructure: better identity, safer agent frameworks, stronger verification and more useful personalization. That does not make foundation models irrelevant. It places them inside a broader system.

This is the deeper significance of his critique. He is not merely dismissing AGI hype. He is proposing a different path to broadly useful intelligence: not one mind, but a society of agents; not one oracle, but a network; not one release date, but a gradual buildout of trust infrastructure.

The current evidence base for this specific public discussion is still narrow. The most detailed fresh account is Forbes’ October 23, 2026 article; other recent sources mostly summarize or reproduce its framing [1], [2], [3], [4]. But the idea itself is substantial because it identifies a bottleneck that model-centric debates often miss. If AI agents are going to act in the world, the central question is not only whether they are intelligent. It is whether they can be found, trusted, limited, tested and held accountable.

Raskar’s answer is a decentralized, verifiable agent ecosystem. If that vision takes hold, the AGI debate may shift from asking when machines become generally intelligent to asking how societies build the roads, credentials and rules that let billions of artificial agents operate without breaking trust.

Sources from the last 72 hours

1. [1]Are We Near AGI? | Daily ArtifactOct 23, 2026, 11:20 PM
2. [2]Are We Near AGI? | CollectorOct 24, 2026, 2:00 AM
3. [3]Unpacking the Myth of AGI: MIT Experts Reveal the Jagged AI FrontierOct 23, 2026, 5:55 PM
4. [4]Are We Near AGI?Oct 23, 2026, 5:55 PM

AI-generated article based on recent web research, then preserved as a dated editorial snapshot.