An agentic L1 is the class of blockchain running AI inference, verifying model output, and settling transactions for autonomous agents, with the inference executed inside consensus.

The category sits apart from a fast chain that agents happen to use, because the model call is a consensus operation rather than a client-side one.

Inference runs on a subcommittee of validators with GPUs attached. Models run at fixed seeds and temperature zero, so every validator independently produces byte-identical output, and the network can agree on what the model said. And the call is metered like any other resource, priced in Somnia’s developer documentation at 0.07 per validator for an LLM inference request against a default subcommittee of three.

Chains that ship all three get applications that arbitrate and settle without leaving the trust boundary.

Chains that ship only the throughput are fast rails for somebody else’s oracle.

• Paul draws the category line at membership rather than usage: “what they’re talking about is agents using their chain rather than agents being part of the chain,” which is the difference between a fast settlement venue and an agentic L1.
• The hard problem was determinism, not throughput. Somnia runs models at temperature zero with fixed seeds, so validators produce identical outputs, because a model that disagrees with itself cannot be put to a vote.
• The prediction market case is an attack on human arbitration. A $160 million Polymarket market on whether Zelenskyy wore a suit resolved through UMA token voting, and Paul calls whale lobbying over Discord “an absolute mad way to do this.”
• Caliber is the most honest part of the design. Paul concedes the resolution layer is abusable, so every market gets a bond-style rating on the reputability of the data sources feeding it.
• The thesis runs well ahead of the numbers. Paul projects $100 billion of agent-driven on-chain volume inside two years while DefiLlama records roughly $2.2 million of TVL across Somnia’s fourteen tracked protocols today.

1. Who Is Paul, and Why Does the Improbable Background Matter?
2. What Does Agentic L1 Actually Mean in 2026?
3. Why Did Somnia Walk Away From Gaming?
4. What Breaks When You Put an LLM Inside Consensus?
5. How Does an LLM Resolve a Market Without a Human Arbiter?
6. Why Does dreamDEX Run Zero Fees, and Who Pays for It?
7. How Big Does Agentic On-Chain Volume Actually Get?
8. What Is Still Missing Before Agents Trade at Scale?
9. How Do You Evaluate an Agentic L1 After This Episode?
10. What Are the Challenges and Limitations?
11. What Are the Regulatory Considerations for an Agentic L1?
12. What Does This Episode Mean for Your Go-to-Market?
13. How Surgence Labs Does This
14. FAQs

Paul spun Somnia out of Improbable as an independent entity and has run it to market since. Before that, he was building crowd technology for entertainment, not financial infrastructure.

The distinction shapes everything that follows. Improbable’s product put tens of thousands of people inside the same digital space at the same time, then tried to monetize them with interoperable digital items. A T-shirt bought at a concert that you could wear inside a game.

Interoperability of that kind needs a trusted global data store, which is what pushed the team toward blockchain in 2020. What they found was that nothing on the market could take the load.

“If you have these experiences where tens of thousands of people show up and they’re all trying to interact on-chain at the same time, blockchain would just blow up. They’re not really designed to do that. They’re designed for slower-moving financial applications.”

Paul is direct about what that origin bought him. “We created Somnia to solve a problem we had, unlike a lot of people that build tech for the sake of building a problem to solve.”

Read that as a claim about sequencing rather than virtue. Teams that build infrastructure first and hunt for demand afterward end up marketing a spec sheet, which is the failure mode this episode keeps circling. A chain with a founding customer, even an internal one, has a shorter path to a use case than a chain with a benchmark.

The term is doing two jobs at once in the market, and separating them is most of the analysis.

Version one is a performance claim. The chain is fast, agents transact quickly, and the agentic label is positioning applied to throughput that already existed. Almost every chain using the word means this.

Version two is an architectural claim. The model call happens inside the protocol, validators reach consensus on its output, and a smart contract can therefore make a decision that depends on unstructured real-world data without trusting an outside service.

Paul is building version two, and he is candid that version one was already solved and then some. “We almost built too far out in terms of performance. Wait two, three years, you’ll need the level of performance Somnia has built for. But right now, today, I don’t actually think the applications exist yet.”

The two versions look identical in a pitch deck and behave nothing alike in production.

Ask which column a chain is in before you ask how fast it is.

Four terms in this episode get used interchangeably in the wider market and should not be.

Confusing the protocol with the app it ships inside is the most common error in this category, and Paul himself stumbles over it mid-episode before correcting.

Names matter here because they set audit scope.

Because the demand was not there, and Paul says so without hedging.

“We’ve moved away from gaming and entertainment because those sectors are quite challenged. I’m long-term bullish on gaming, but at the moment it’s not doing very well in Web2 and Web3.”

The replacement thesis has two legs. Agents were becoming good enough to delegate to, which Paul dates to Q3 of last year on the engineering side, and DeFi was the one on-chain category with demonstrable product-market fit. Somnia announced the completed repositioning on 21 April 2026, citing more than 2 billion transactions processed across six months and 80 million in a single day.

Grant the counterargument its full weight. Pivoting a chain’s category after a token launch is expensive; it strands the developers who built for the old thesis, and it invites the reading that the original market never showed up.

Paul takes that hit directly rather than dressing it. The games are still in the ecosystem; they are not the story anymore, and the honest version of the repositioning is that a chain built for real-time entertainment found its buyer somewhere else.

The applied lesson for anyone running a similar move is that the pivot only survives if the new claim is architectural. Repositioning from “fast chain for games” to “fast chain for agents” is a slide deck. Repositioning to a chain that executes inference under consensus is a different product, and it can be verified.

Positioning that cannot be checked gets discounted to zero by the people you most want reading it.

Start with why anyone would want to. Paul frames it as a decentralization problem rather than a capability one.

“If you want to make a fully decentralized application and you want to use agents, you have to call out to a central service like OpenAI’s or Anthropic’s. These are centralized control points that mean the app is no longer decentralized.”

So the goal is a contract that can reason over messy real-world input without importing a single point of failure. Getting there broke four things.

Determinism. Two validators running the same prompt against the same model will normally return different text, and a network cannot vote on output it cannot reproduce. Paul was surprised this was solvable. “I didn’t think you could run LLMs in deterministic modes, but you can. It’s brand new research.” Somnia’s fix is fixed seeds and temperature zero so every validator produces byte-identical output. Cost of getting it wrong: inference calls fail non-deterministically and every application built on them is unreliable in production.

Validator hardware. Consensus now requires GPU-equipped compute nodes inside the validator set. Cost of getting it wrong: the validator set concentrates around whoever can afford the hardware, and the decentralization argument that justified the whole design collapses.

Metering. An LLM call is orders of magnitude more expensive than an EVM opcode and needs its own gas model. Somnia’s documentation prices a JSON API call at 0.03 per validator, an LLM inference at 0.07, and a website parse at 0.10, with a default subcommittee of three. Cost of getting it wrong: either inference is free and the network is trivially spammed, or it is priced like compute and nobody builds on it.

Latency. Inference is asynchronous. The contract calls createRequest(), gets a request ID back immediately, and receives the answer later through a handleResponse() callback that can return Success, Failed, or TimedOut. Cost of getting it wrong: developers write contracts that assume a synchronous answer and discover the failure path in production.

Note what the pricing table implies. Every inference is paid for three times, once per validator in the subcommittee, and that multiple is the price of the trust guarantee.

Verification is never free. It is only ever cheaper than the alternative.

Building on an agentic L1 and struggling to explain the architecture to anyone outside engineering? Book a free 30-minute go-to-market call. No pitch deck. No obligation.

This is the sharpest section of the episode, because Paul is not describing a feature. He is describing an attack on how the incumbents work.

Most large prediction markets resolve through UMA, an optimistic oracle where token holders vote on contested outcomes. Paul has watched what that produces.

“If you look at the big Polymarket markets, they’re actually massively lobbied by whales, and they get changed by these whales chatting to people and influencing people on Discord, which is an absolute mad way to do this, in my view.”

The reference case is public. A market on whether Volodymyr Zelenskyy would wear a suit before July drew more than $160 million in volume and resolved against what many participants saw at the NATO summit, with the “yes” price collapsing from 19 cents to 4 cents. Critics argued that a large enough block of UMA holders can vote a result into existence.

Prophecy replaces the vote with a model. A market creator writes a prompt, the LLM identifies which sources should decide the question, and the resolution path is rendered as a flowchart before anyone can take a position.

Paul grades the difficulty honestly. Sports results and weather are straightforward because the public sources are dense. Subjective questions are not, and his answer there is breadth rather than certainty: pull news, X posts, Reddit threads, and a wide spread of sources, then let the model synthesize a call.

He is equally clear about the hole in it. “You can abuse this. I could set up poll.com, and that could become a data source.”

The answer is a rating layer. Caliber scores each market on the reputability of the sources feeding it, on a bond-style scale from AAA down through B, and every market carries one.

That is the design decision worth copying. The system does not claim the model is always right. It claims the inputs are legible and graded, and it puts that grade in front of the trader before capital moves.

A resolution system that hides its sources is a governance vote with extra steps.

dreamDEX is the first-party proof point, and the strategy behind it is stated openly. “The problem with new technologies is you have to get them adopted. A lot of people build the infrastructure and just assume people come and build, but you actually have to drive people with use cases and examples.”

The venue makes three claims, and each has a cost attached that the episode names.

Everything on-chain. Matching, liquidation, ADL, and fee logic all execute on-chain. Paul is scathing about the alternative. “Most of them are what I call faux DeFi. They run the matching engine, the liquidation engine, everything off-chain, and they just have deposits on-chain.” His stated motivation is personal: “I’ve now been in the space on the other side of making a token, and I’ve realized there are lots of bad things with exchanges.” On 10 October he was on the losing side of the $19 billion liquidation cascade and says so.

Latency competitive with TradFi. Somnia currently runs 100-millisecond block times with sub-second finality and a peak recorded 134,642 transactions per second. New technology in the pipeline targets sub-10-millisecond blocks. Paul is precise about who benefits: not the retail click-trader, but the institutional market maker who can quote tighter when latency drops, which reaches retail as spread.

Zero maker and zero taker fees, permanently. The dreamDEX positioning post describes a yield-based CLOB where stablecoin yield compensates market makers instead of trading fees. Every pair is denominated in USDSO, Somnia’s Frax-backed native stablecoin, and the yield those balances generate funds the maker incentives. Paul is blunt about the business logic: “There is gas. If we’re running a blockchain business, we’re trying to drive usage back to the blockchain.”

The counterparty question that follows is the right one to ask. Two competitors already run zero fees, and Paul argues both pay for it somewhere the user cannot see, one through slippage and one through preferential treatment for institutional makers.

Which is the test to apply to dreamDEX in twelve months. Zero fees always get funded by something. Find the something.

A venue that cannot tell you where its revenue comes from is telling you anyway.

Ruthy anchored the question at roughly $10 billion of daily DEX volume and asked whether agents move that number.

Paul gave two paths and a timeline. Agents transacting on-chain drive an order of magnitude, taking the market to $100 billion inside two years and trillions over five to ten. Traditional finance running on the same rails does the rest, and he points at perpetuals venues starting to compete with incumbent exchanges as the leading indicator. The same institutional path that ETFs opened for spot exposure is now being cut for execution.

He then undercuts his own enthusiasm, which is the most useful thing he says about timing. “AI has been cooking for 30 years until it’s got to here, so these things take time. Blockchain has not been around that long.”

That is a claim about where the cycle sits rather than about the technology, and it is the more useful of the two.

The current numbers support the direction and not the pace.

The middle two rows are the ones to hold together, and the top row deserves the same scepticism you would apply to any volume print you did not measure yourself. Agents are transacting constantly and settling almost nothing, which means the volume case rests entirely on average transaction value climbing, not on transaction count.

Count is the vanity metric here. Value is the business.

Paul names trust as the binding constraint, and he includes himself in the diagnosis. “This is going to take time. It’s a trust function. I’m probably right on the bleeding edge.”

Three gaps sit underneath that.

Delegation depth. Agents currently execute instructions well. Handing one discretionary authority over a portfolio is a different decision, and Paul frames the pitch around emotional discipline rather than alpha: traders blow up because they double down after a loss, and a model does not. It is the same behavioral argument that made copy trading a growth engine for exchanges, applied to a counterparty that has no ego. Persuasive, and still unproven at size.

Liquidity for the long tail. Prophecy lets anyone create a market, which means most created markets have no counterparty. Paul concedes it plainly with a hot dog example, and the protocol market maker only seeds markets that show interest. Permissionless creation produces a graveyard by default.

Yield sources that survive a cycle. dreamDEX plans yield-bearing collateral, routing unused margin into lending protocols automatically, and Paul is looking at real-world assets paying 8 to 12 percent to extend it. The distinction he draws between a rate that comes from borrowers and a rate that comes from emissions is the same one that separates durable DeFi yield from a TVL rental agreement. He rules out the shortcut: “We don’t have plans to juice these yields up with massive token emissions. Those work for a while and you get more TVL, but then it all disappears.”

Get those three right and the delegation curve compounds. Get them wrong and an agentic L1 is a very fast chain with idle capacity, which is precisely what Paul admits Somnia was before January.

Treat the label as a claim to be tested rather than a category to be accepted. The episode supplies the questions.

Paul describes dreamDEX’s rollout in enough detail to reconstruct the sequence, and it is a defensible template for launching a venue on a new base layer.

• Private alpha. Perpetuals are already running internally at the time of recording. Nothing public, no incentives, risk engine under observation.
• Invite-only launch. Access codes gate entry. The stated reason is risk management rather than exclusivity, since “a perp DEX is a very complex beast.”
• Volume competition first. Algo Arena runs roughly $10,000 in prizes with a weekly leaderboard reset, deliberately weighted toward algorithmic traders over click traders.
• Quality gates before scale. Paul names explicit thresholds that must be cleared before the incentive program opens, because a points campaign drives usage into whatever is behind it.
• Points program last. Prize pools in the $50,000 to $100,000 range, weighted toward stablecoins rather than token emissions. His reasoning: “The problem with token-based pools is maybe the token goes bananas and everyone makes tonnes of money, but maybe not.”
• Parallel private track. A Founding Traders program invites from a list of the top 20,000 on-chain traders, running alongside the public funnel rather than after it.

The ordering is the lesson. Incentives come after the risk engine has been watched under real flow, not before, which is the opposite of how most launches sequence it.

Points before quality gates is how a venue buys volume it cannot survive.

Launching a venue or a chain into a category the market has no word for yet? Book a free 30-minute go-to-market call. No pitch deck. No obligation.

The interview is a founder describing his own architecture, and four things go unresolved.

Determinism is not truth. Fixed seeds and temperature zero guarantee validators agree on the output. They guarantee nothing about whether the output is correct. A model that reads its sources wrong reads them identically wrong across the whole subcommittee, and consensus certifies the error.

Source selection is the new attack surface. Prophecy moves the manipulation point from the vote to the input list. Caliber grades that list, which helps, and a grading system is itself a target once enough money depends on its output.

The model is a dependency with a version number. Somnia does not control the weights it runs. A model upgrade changes behaviour across every contract that calls it, and the episode does not address how versions are pinned or migrated.

The gap between claim and balance sheet is wide. Peak throughput of 134,642 transactions per second against roughly $2.2 million of chain TVL is a large ratio, and no amount of architecture argument closes it. Only applications do.

• Marketing throughput as an agentic claim when the model call runs off-chain.
• Treating deterministic inference as verified inference.
• Launching permissionless market creation with no liquidity policy for the long tail.
• Opening a points program before the risk engine has been tested against real flow.
• Denominating everything in a native stablecoin without stating who holds the reserves.
• Naming a protocol and its flagship application the same thing, then wondering why the audit scope is unclear.
• Quoting peak TPS from a test window as though it were sustained capacity.

Two regimes converge on this architecture, and neither was written with it in mind.

The first is prediction market oversight. On 12 June 2026, the CFTC issued a proposed rule on public interest determinations for event contracts, setting out when contracts may be prohibited as contrary to the public interest and defining gaming as activity participants “typically engage in for purposes of recreation or to entertain others.” The comment period ran to 27 July 2026. A protocol whose selling point is that anyone can create any market permissionlessly is describing the exact listing discretion this rulemaking is trying to constrain.

The second is the EU AI Act. Article 26 places obligations on deployers of high-risk systems, including the requirement that they “monitor the operation of the high-risk AI system on the basis of the instructions for use” and assign human oversight to people with the competence and authority to exercise it. Those obligations apply from 2 August 2026.

Hold those two facts against the design. An LLM that decides whether an insurance policy pays out is making a determination about a person’s financial entitlement, and the deployer of that system carries the monitoring duty regardless of whether the inference happened inside consensus.

Decentralized execution does not decentralize liability. It only makes the deployer harder to find, which regulators read as a problem rather than a feature.

The practical exposure sits with whoever fronts the application. Publish the resolution prompt and the source list, keep the Caliber grade visible before capital moves, retain the callback records, and name a human accountable for oversight of the deployed system. Treat it the way a serious team treats reporting obligations, as a standing process rather than a filing exercise.

Build that documentation while the product is small. Retrofitting an audit trail after a disputed payout is the most expensive version of this job.

Three things carry over to any project shipping in this category.

A category claim has to be checkable. Paul’s whole differentiation rests on one sentence a technical buyer can verify in the documentation: inference executes on validators under consensus. That is a stronger position than any adjective, and it is why the repositioning reads as substance rather than a rename. If your category claim cannot be checked against a repository or a spec, it is a tagline.

Ship the first-party application yourself. Somnia built Prophecy Social and dreamDEX rather than waiting for developers, on the explicit theory that infrastructure without a worked example does not get adopted. This is the same discipline that separates a token launch with real liquidity from one with a chart nobody wants to catch. Demand has to be demonstrated before it can be delegated.

Sequence incentives behind proof. The rollout puts a small competition before a large one and quality gates before the points program. Most teams invert this because the incentive budget is the easiest thing to spend. Run it the other way, and the campaign lands on a product that can absorb it. Run it the usual way and the campaign is the stress test.

The pattern holds beyond agentic infrastructure. It applies to any exchange or venue trying to win its first thousand traders, to cross-chain expansion decisions where each new deployment becomes its own funnel, and to how you brief the KOLs and agencies carrying the claim on your behalf.

An audience recruited by incentives leaves when the incentives do.

We run go-to-market for projects launching into categories the market has not named yet, which is the exact problem an agentic L1 has. The work is positioning that survives a technical audience, narrative that a non-technical buyer can repeat, and campaign sequencing that puts proof before spend.

That means writing the checkable claim rather than the adjective, building the trader and developer funnels separately because they convert on different evidence, and holding incentive budget back until the product can take the flow.

We have run this for Binance, Animoca Brands, Polygon, Falcon Finance, Incentiv, Rails, Biptap, SendIt, Black Mirror, and Boost.

Bringing a new category to market and unsure whether the claim survives contact with a technical buyer? Book a free 30-minute go-to-market call. No pitch deck. No obligation.

The useful test this episode leaves you with is a single question. Does the model call happen inside the protocol, or does the chain simply move fast enough that agents like using it?

Answer that and the agentic L1 category sorts itself into two piles quickly. One pile is positioning applied to throughput. The other is a genuinely different trust boundary, with determinism, validator hardware, metering, and an asynchronous failure path all working, and with the honest caveat that identical output is not correct output.

Somnia is in the second pile on architecture and near the bottom of the first on adoption, and Paul does not pretend otherwise. “Even if the bull market doesn’t come back, we’re in this for the long term. We’re going to build this for the next five, ten years.”

The chains that matter in this category will be the ones whose applications force the inference layer to be used, not the ones whose benchmarks say it could be.

It covers Somnia’s repositioning as an agentic L1, the engineering behind running AI inference inside consensus, and the two applications built on it, Prophecy Social and dreamDEX. Ruthy hosts. The episode runs 57 minutes and was recorded in August 2026.

An agentic L1 is a base layer that executes and verifies AI inference as part of block production, so a smart contract can act on unstructured real-world data without calling an outside service. The distinguishing test is whether the model call is a protocol operation or an application one. Most chains using the term mean the second.

A contract calls createRequest() and receives a request ID immediately. Execution happens off the EVM on a subcommittee of validators with GPU hardware, and the result returns asynchronously through a handleResponse() callback that reports Success, Failed, or TimedOut.

By making the output reproducible. Somnia’s documentation states that agents run with fixed seeds and temperature zero so every validator independently produces byte-identical output, which is what makes consensus on a model result possible at all.

No. Determinism guarantees that every validator produces the same output, not that the output is right. A model that misreads its sources will misread them identically across the subcommittee, and consensus will certify the error.

Somnia’s developer documentation prices a JSON API call at 0.03 per validator, an LLM inference at 0.07, and a website parse at 0.10. With a default subcommittee of three validators, the cost is paid three times per request.

Prophecy is the protocol that lets anyone define, seed, and resolve a prediction market on-chain. Prophecy Social is the first-party application built on it, running social markets with no real money as a proof point for the protocol.

UMA settles disputes through token-holder voting, which Paul argues is open to whale lobbying, citing markets changed by influence rather than facts. Prophecy resolves through a model reading a published list of sources, with the resolution logic shown before the market opens. The manipulation risk moves from the vote to the source list.

Caliber is Somnia’s rating system for prediction markets. It scores each market on the reputability of its data sources using a bond-style scale from AAA down through B, so a trader can judge resolution reliability before taking a position.

Trading is zero maker and zero taker fee, and the venue earns from gas paid on Somnia plus yield generated on USDSO balances held on the exchange. That yield currently funds market maker incentives rather than the treasury, which is a growth-phase choice rather than a permanent model.

Only if the application genuinely needs a decision made on unstructured data inside the trust boundary, such as arbitration, claims, or resolution. If the agent is only submitting transactions, a fast chain with mature liquidity will serve you better and cost less to integrate.

Related Reading

• Onchain AI Agents: Inside the Surgecast Episode With KeeperHub’s Simon
• What Is Liquidity in Crypto and Why Does It Matter for Your Token Launch in 2026?
• Crypto Exchange Marketing: How to Win Your Next 1,000 Traders
• DeFi Yield Farming Explained: What Crypto Projects Should Know in 2026
• Blockchain Interoperability: Why Cross-Chain Matters for Crypto Project Growth in 2026