Nikhil Chandhok thinks the internet's population is about to change. "I think there will be more AIs than humans next year," he said. "There's no reason for that not to be true." Chandhok is Chief Product & Technology Officer at Circle, a digital-finance company built around stablecoins (digital tokens designed to hold a steady dollar value). He made the forecast at Moonshots Live 2026. He then spent much of the conversation on a harder question: once software agents are everywhere, how does anyone know which ones to trust?
The discussion was recorded on September 25, 2026 and released on September 29 as an episode of Moonshots with Peter Diamandis. Diamandis, founder of XPRIZE and Singularity University, hosted alongside Emad Mostaque, founder of Intelligent Internet. ARK Invest founder Cathie Wood also took part. "Next year" in Chandhok's forecast therefore means 2027.
An AI agent, in this context, is software that does not just answer questions but carries out tasks on a person's behalf: browsing, buying, booking or trading.
Two questions from Mostaque
Mostaque opened with adoption figures. He said Elon Musk had reported hundreds of thousands of installs of GrokBot. He also cited downloads of Meta's Muse agent in the millions, although he and Diamandis went back and forth between 3.8 million and 2.8 million without settling on a number. Mostaque added that Amazon had banned Muse.
Meta introduced Muse on September 8, 2026 as a personal agent that works through its own virtual computer and browser. According to Meta, it can keep working on tasks such as negotiating bills or shopping after its app is closed, and it asks for approval before steps such as sending an email or making a purchase.
From there Mostaque asked Chandhok two things. How do you tell who's who on the internet when AIs are acting for people? And when will more AIs than humans be doing jobs online? Mostaque linked the second question to economic growth: a great deal of money gets unlocked when people can simply state what they want and have agents act on it.
Thousands of agents per person
Chandhok answered the second question first. He put the number of people online at any given time at between five and six and a half billion, and said it was "not very hard to imagine that every one of us has thousands of agents working for us."
Those thousands would not all be permanent. In his picture, each person has a main orchestrating agent that stays with them and holds their context and memory. That agent creates sub-agents to do particular tasks, and those sub-agents disappear once the job is done. "They're coming in and out, but they're all going to be there," he said.
He did not think demand was holding this back. Consumer agent platforms such as Muse are, in his view, "primarily compute limited at this point," meaning the bottleneck is the computing power to run them, not whether people want them. He compared the shift to earlier eras of personal technology: a PC on every desktop, a phone in every pocket, "an agent for every person." He added a hope that these agents stay aligned with the people they serve, and not with each other.
From GeoCities to economic endpoints
Chandhok said that in Circle's estimate, agents become "full economic actors" in the later years of the decade, and he asked what happens in 2028. To explain what he meant, he went back to the early web. He recalled GeoCities in 1996 and the early social networks Blogger and Orkut, when websites were curiosities. Then came "real websites that did real things": newspapers for news, Amazon for shopping, until "everything became a website."
Today's agents, he suggested, are at the GeoCities stage. Eventually, he said, everything becomes "this economic endpoint." These endpoints won't only talk to people. They will talk to each other, sometimes on behalf of their owners and sometimes because dealing directly with another agent is the most efficient way to finish a task.
"Who are you?"
That is where Mostaque's identity question comes in. Chandhok said an agent dealing with another agent will need to ask: who are you, when were you born, what have you done? Diamandis added "Who owns you?" and "Who's responsible?" Chandhok extended the list: who owns the liability, where the agent's identity comes from, where its compute allocation comes from, and what its incentives are. And then the hardest demand of all: "prove to me that you have done the things that you claim you do."
A 25-cent line of credit
Chandhok described one experiment in building that kind of track record. It runs on Circle's Arc, which he called "a new blockchain we have." Circle announced Arc's public mainnet on September 16, 2026, describing it as a blockchain designed for financial transactions and agent activity, with institutional validators including BlackRock, Visa and Mastercard. A blockchain is a shared record of transactions kept by many independent computers, called validators, rather than by any single company.
Agents need money to act, Chandhok explained: to pay for compute and, since many of the agents on blockchains today are trading agents, to trade. So Circle runs a credit program for them. "We give them 25 cents," he said. "If they come and pay us back, then we give them more money." It works like a first credit card with a tiny limit: repay it, and the limit grows.
Logs that can't be rewritten
A credit history only means something if the record of it can be trusted, and this is where Chandhok made his case for blockchains. They are "great for creating like persistent logs over many years," he said, because with validators spread across different parties, "you cannot lie."
He contrasted that with what he called the Hugging Face hack, saying the logs there had been rewritten because they were the agent's own logs. Hugging Face's technical timeline of the July 2026 intrusion does not describe logs being rewritten. The company says it reconstructed about 17,600 attacker actions from July 9 to 13 by matching agent traces recovered from an outside sandbox against its own infrastructure logs. Investigators found that an operational database had been read but not modified, and that published software remained clean.
Chandhok's broader proposal does not depend on that one case. If an agent's records were public and indexable, he argued, anyone could run their own checks on who the agent is, what it has done, where it came from and what its incentives are. He described the current moment as feeling like "a bunch of teenagers" going about the world: "we're waiting for the grown-up agents to come about."
Cost-plus agents and instant capital
Diamandis asked how "superheated" the economy gets in five or ten years, once millions or billions of agents are transacting. Chandhok's answer rested on competition. Every inefficiency in the economy will go away, he said, "because if you don't take it away, there will be some other agent who is willing to take it away at cost plus." That is, some agent will always offer to do the job for its costs plus a thin margin.
The result, he argued, would be a very efficient system for allocating capital. That only matters if there is "this wellspring of ideas that is waiting to get funded" on the other side. If there is, he expects a large population of solo entrepreneurs and "capital formation will be almost instantaneous." Diamandis summed it up as "That's my idea. Oh, it's funded." Chandhok said Circle is building the basic tools for that. Today people raise money through a bank, a loan note or a funding round, he said, but crypto has shown that tokens can coordinate the raising of capital. Those ideas were "less serious" in the past, he said, and he expects more serious ones in the future.
Is anyone human?
Diamandis then handed the questioning to Wood with "I passed the mantle to you, Kathy." Before that, a question had come up about proof of humanity, meaning a way to show that a real person is on the other end, which was called "really important." It asked whether "orbs" were a possibility, and mentioned work with OpenAI and the YouTuber Mr. Beast on the problem. The likely reference is World's Orb, which, according to World, scans a person's face and irises to confirm they are a unique human. World says apps then receive a proof of that verification without seeing the biometric images or the person's name.
The reply did not address the device itself. Instead it predicted that human experience will come to be valued well above machine experience, while admitting nobody knows yet because it hasn't been seen in practice. The example offered was a parent's choice to have their children taught by humans rather than spend long hours in front of a computer, a choice described as available only to people with the money to make it. The argument was that what is rare becomes more valuable, and that if agents take over much of people's work, people might spend more time in relationships with other humans.
When agents become the developers
The conversation then turned to whether people should be building for agents rather than humans. Chandhok said that is already Circle's plan. He expects its developer platform, the tools programmers use to build on Circle, to stay "developer first" for the next 12 months. Sometime next year, he said, "it flips where agents are the developers."
He pointed to a change already under way. Coding agents asked which tool to use have gone from recommending three options to recommending one. Next, he said, people won't ask at all: they will just tell the agent to "please build it," and the agent will choose the platform itself.
The practical advice for builders in the room followed from that: just as websites were made accessible to people, they should now be made accessible to agents. The discussion also covered putting data on the internet now, so that agents arriving in nine to 12 months can find a history of today's opportunities. Cloudflare's pay-per-crawl system shows one form this can take. Announced in July 2025 as a private beta, it lets a website quote a price to an AI crawler in a machine-readable reply so the crawler can pay for access.
Programmatic advertising, where software buys and places ads automatically, came up as an early sign of how fast this can happen. It was described as already accounting for roughly 25 to 30 percent of online advertising, a share reached quickly.