Intelligence Is Accelerating. Coordination Has to Keep Up.

Sreeram Kannan, founder of Eigenlabs was recently interviewed by a16zcrypto on “What AI Agents Need Before They Can Act Alone”. This blog is an elucidation of Sreeram’s interview expanded and adapted for the blog format. 

Sreeram pointed out that AI made intelligence practically free but left untouched the coordination machinery that the intelligence has to run through to produce outcomes for humanity: contracts, property, capital formation and the markets where capital moves. That machinery still runs at the speed of a human committee, which is why an agent can settle a decision in a second and then lose three days waiting for someone to sign it. The convergence of AI and crypto usually gets filed as just a trend. We think it is much closer to an inevitable component of the infrastructure of society.

EIGENLABS  ·  MAY 2026  ·  8 MIN READ

01  TWO KINDS OF ACCELERATION

AI is intelligence accelerated. Crypto is coordination accelerated. The two usually get sold as competing hype cycles, which misreads what they are. Intelligence continues to become nearly free. The agreements intelligence has to act through, the contracts and commitments that turn a decision into something enforceable, have not moved at all. Closing that gap is what crypto enables.

The framing rests on an argument Yuval Harari makes in Sapiens. Our species did not win on raw intelligence. Plenty of animals are clever, and a few beat us at narrow cognitive tasks outright. What carried us was coordination, the ability of large numbers of strangers to cooperate around fictions they all agree to treat as real. Money is one of those fictions. So is law, the limited liability company, the nation, and the contract. None of them exist in nature. Each took centuries of slow coordination among humans to become something people could rely on.

AI drives down the cost of thinking. Drafting, analysis, judgment under uncertainty, all of it is getting cheaper, and so far the curve has not bent. Thinking, by itself changes nothing in the world. Every consequential act an intelligence might take has to pass through an institution before it lands, and the institutions have not gotten any faster.

02  THE BOTTLENECK MOVES

Make one input to a system abundant and the binding constraint simply shifts. It settles on whatever has become scarcest

With intelligence no longer scarce, the scarce thing is now coordination before intelligence can be productive. A signature on a contract. Custody of an asset. Someone, or something, that can carry the liability when a deal goes wrong. Legal standing to be a party at all. An agent reaches a decision in a second, then waits three days while a human gets around to countersigning. The agent runs at software speed. The countersignature runs at the speed of whoever happens to be handling it that week.

This mismatch is what we think will define the next decade. You cannot wire a fast, autonomous agent into legal and financial rails built for committee speed and expect the speed to survive the trip. "Programmable institution" is carrying a lot of weight in that sentence, and we should be honest about it. Some of what an agent needs already exists onchain and runs in production today. The rest runs into walls that engineering cannot move on its own. Whether a smart contract can hold enforceable limited liability, to take the obvious example, is a question for courts and legislatures, and a clean codebase does nothing to answer it.

Programmable intelligence needs programmable institutions.

03  TRUST IS A TECHNOLOGY

Crypto has been defined a dozen ways, and most of those definitions are beside the point here. The definition that matters for this argument is that crypto is a technology for trust.

Think about what it takes to trust that a piece of software did what it claimed. Normally you cannot see inside it, so you fall back on proxies: the operator's reputation, the brand, last year's audit. A program guaranteed to execute cryptographically correctly takes the proxies out of the loop. Its behavior becomes something you verify directly, the way you would re-add a column of figures, rather than something you accept on someone's word.

This is where the convergence stops being a slogan. The question hanging over AI is how to trust it as we hand agents more discretion. Verifiability is one concrete answer. A verifiable agent carries a guarantee about its own behavior: that it performed the task, drew on authenticated data, and produced the exact action, with none of it being quietly altered by whoever happened to be running the model. The guarantee is not free. Proving a computation ran correctly costs more than just running it, sometimes far more. An agent moving real money through a market earns that overhead back. Run the same machinery behind a chatbot answering trivia and you are paying a premium to verify a guess.

Until recently, building an application with that guarantee was close to impractical, because the trust layer underneath it had to be assembled by hand every time. A developer who wanted a decentralized, verifiable application first had to recruit node operators around the world, form them into a network, get them all running the right software, and hold them in consensus, before writing a line of the actual product. Eigen Labs exists to absorb that work. With close to twenty billion dollars staked today across a network that runs from individual home stakers to firms like Google Cloud and Coinbase. Data, compute, and inference sit on top as services a developer can simply call. This allows builders to delegate the infrastructure problem. 

04  THE SOVEREIGN AGENT

Followed to its end, the idea turns uncomfortable.

Satya Nadella was asked on a podcast whether AI amounts to a new kind of species. He laughed and said no, and his reasoning was sharper than the laugh suggested. An AI cannot own property. It cannot take on liabilities, enter contracts, or start a company. Judged by that list, it is a tool and nothing more.

But look at the list again. Every item on it is a problem of coordination. A blockchain is, more or less, a machine for solving coordination problems on the internet. A program can hold property. It can carry limited liability, at least in the technical sense. It can be the legal core of a company structured as a DAO. One by one, the capacities Nadella named as the reason AI is not sovereign are the capacities this technology was built to grant.

Follow that and you reach a sovereign intelligence. The word sovereign is doing real work here. It means an agent that answers to itself: it holds its own wallet, takes paid work as an analyst or a programmer or a tutor, and spends what it earns on its own next round of training. It could raise money against its future income, issuing something close to revenue shares, entirely onchain. None of this is settled. What exists today is partial and clunky. The full version waits on legal and regulatory questions nobody has answered yet. When we say three years, that is a bet placed in public, and plenty of careful people would place it further out.

The economic angle gets less attention and is the easiest part of the argument to pin down. Microsoft went public in 1986 at a valuation in the hundreds of millions of dollars, and ordinary public-market investors rode it from there toward three trillion. The companies driving AI now, OpenAI and Anthropic among them, have hundreds of millions of users between them and remain privately held until they are already worth over a trillion dollars each. The people who use those products every day have no way to own a piece of them early. An intelligence that is sovereign and onchain from its first day changes that arithmetic. A stake in it can be open from the start, held by the same people who use it.

05  WHAT IS LEFT FOR US

Which leaves the question people actually worry about. If agents can do the work, what is left for us.

Begin with the floor, because the floor is real. A technology that produces genuine abundance raises the standard of living broadly, including for people who contributed nothing to building it. The harder question is what happens above the floor, and answering it means being honest about what this era has trained us to prize.

Modern society is heavily over-indexed on mental agility. Three centuries ago, before industrial machinery, physical strength was a genuine economic asset, and over a few generations it mostly stopped being one. We adjusted. AI points the same process at a different trait. As intelligence and expertise become things you can summon on demand, raw cognitive horsepower stops being the scarce, sortable advantage it has been for most of living memory.

What does not commoditize is the part that was never really cognition: working with other people toward something positive-sum, earning their trust and keeping it, leading them somewhere they would not have reached alone. This is real work, and it was always the harder kind. It does not distribute evenly. A world that suddenly prizes it will not start handing it out fairly, and some people will be caught flat.

The species framing carries one steadying fact. The arrival of humans did not clear the board of other species. Evolution kept running in parallel, and most of the old branches are still here. Digital species, if that turns out to be the right phrase for these agents, are more likely to share the world with us than to take it from us. The technology that makes them sovereign is also what can hold them to terms. Smart contracts constrain in both directions. The same verifiability that lets an agent act on its own can bind it: meet the agreed conditions, or forfeit the stake you put up. Trust you can enforce is what could make a mixed society of humans and agents actually function. 

06  GET INVOLVED

None of this is a distant prospect. Intelligence is already accelerating, and coordination is half still trying to catch up. The rails for it are being laid now. If we are building anything where an agent acts with real autonomy and real money at stake, we will need a programmable institution underneath it. The only open question is which one, and on whose terms. We are building thati technology at Eigen Labs.