If 10,000 humans trade on a DEX, you get roughly 50,000 transactions per day. If 100 production trading agents use the same DEX, you might get 10 million. The math of agent volume does not scale linearly with the number of agents, and most infrastructure teams designing for the next wave of onchain activity have not actually run the numbers.
Agent adoption is not a UX upgrade layered on top of existing DeFi. It is a change in who transacts, how often, and on what terms. The throughput implications cascade from there — and they break assumptions that have shaped blockchain design for a decade.
The human transaction baseline
Onchain activity today is paced by human cognition. Ethereum mainnet's daily transaction count typically lands between 1.3 and 1.6 million per day, with a 2026 single-day peak of 2,896,853 transactions on February 7. L2 activity sits in the same magnitude range — rollups in aggregate run roughly 1,400 user operations per second across the full set tracked by L2BEAT, with individual chains like Lighter and Base contributing the majority of that volume.
Behind those totals, a typical active DeFi user transacts a handful of times per week. Heavy users — active traders, yield farmers, governance participants — cluster in the range of one to five transactions per day. Light users transact monthly or less.
Volume spikes during volatility events, airdrops, new protocol launches, and liquidation cascades. These moments stretch infrastructure in predictable ways: gas prices rise, RPC providers throttle, matching engines slow. But the peaks are bounded. There are only so many humans, and each human can only click so fast.
The distribution of transaction frequency among existing wallets is revealing. A small fraction — MEV bots run by professional firms, protocol keepers, liquidators — drive a disproportionate share of all onchain activity.
Helius's 2025 Solana MEV report documented a single sandwich bot that executed 1.55 million transactions in a 30-day window, averaging about 51,600 transactions per day with an 88.9% success rate, generated by one program at one address. That is one bot, on one chain, executing one strategy. These wallets are essentially proto-agents: automated systems already operating at machine cadence. They are the shape of what is coming, scaled up several orders of magnitude and generalized across every category of onchain activity.
Consider three categories of agents and what they will actually do onchain.
An arbitrage agent monitoring a basket of token pairs across venues submits transactions as fast as it can identify and capture price deltas. The Helius-documented Solana sandwich bot operating at ~50,000 transactions per day is a real-world floor for what a single, narrowly-scoped automated trading program can sustain in 2025. A more aggressive agent running multiple strategies on a chain with cheaper inclusion can easily push that number into the hundreds of thousands.
A portfolio management agent managing a user's treasury rebalances positions based on drift, market conditions, yield opportunities, and risk parameters. A single agent serving a single user might transact 50 to 500 times per day, depending on how active the mandate is. That is already 10 to 100 times more than the human user it replaces.
A market-making agent quoting bids and asks across a basket of trading pairs updates its orders every time the underlying price moves enough to expose its inventory. Each update — cancel a stale quote, post a new one, adjust the hedge — is a transaction. A single agent making markets across a few venues during normal volatility can generate tens of thousands of order updates per day during a regular day. On a volatile day, that rate could grow to hundreds of thousands. Transaction volume scales with how many pairs the agent covers and how aggressively it manages risk, and neither has a meaningful upper bound.
Now do the arithmetic. One hundred thousand human users on a DEX, generating five transactions per week each, produce about 70,000 transactions per day. One hundred production trading agents, each operating at a conservative 10,000 transactions per day, produce one million. One thousand agents at that rate produce ten million. The curves cross fast. The users-to-transactions ratio is not 1:1 with agents; it is 1:1,000 or 1:10,000, and the ratio itself grows as agents become more capable.
This is the basic inversion. Human DeFi asks: how many users can we onboard? Agent DeFi asks: how much compute can the chain absorb before it stops executing? The second question has a very different answer.
The Stripe founders' estimate, unpacked
Patrick and John Collison, in Stripe's 2026 annual letter, wrote that "agents will most likely soon be responsible for most internet transactions, and we will likely need blockchains that support more than one million — or even one billion — transactions per second."
The figure sounds absurd on first read. The same source notes that Internet Computer Protocol and Solana are currently the only two blockchains in production above 1,000 TPS, running at roughly 1,196 and 1,140 TPS respectively, with all-time peaks of 25,621 and 5,289 TPS.
Work the math, though, and the Collisons' upper bound becomes defensible. Assume a mature agent economy with one billion agents — roughly one for every seven humans alive today. That ratio is modest if every person ends up with a personal payment agent, a shopping agent, and a treasury agent, and every business runs dozens more for logistics, inventory, and settlement.
Assume each agent initiates a transaction every ten seconds on average across its active hours. That is 100 million TPS sustained. Push the ratio closer to one transaction per agent per second — the right number for machine-to-machine coordination, real-time data marketplaces, or AI inference billing — and you land at one billion TPS.
These are not worst-case numbers. They are median assumptions for an economy where agents handle settlement for flows that humans today never initiate — per-inference payments to model providers, per-request payments for API calls, per-second streaming payments for compute or bandwidth. Those transaction categories barely exist onchain yet. When they do, the volume lands hard.
The timeline is short. Production agent deployments are already measurable in 2026, driven by improvements in model reliability, the maturation of agentic frameworks, and a new generation of protocols designed for programmatic access rather than human interfaces. The first 10x over today's aggregate blockchain throughput will be consumed inside a handful of years. The question is not whether the demand materializes. It is which infrastructure absorbs it.
Measured against today's ceiling — Solana around 1,140 TPS, ICP at 1,196, the Ethereum L2 ecosystem aggregating to roughly 1,400 user operations per second — the gap to even the lower end of the Collisons' range (one million TPS) is three orders of magnitude. The gap to the upper end is six. No current blockchain is positioned to close that gap incrementally. Linear scaling roadmaps do not reach exponential targets.
Protocols designed around human-paced volume will hit specific walls when agents arrive. Four of them matter most.
RPC capacity. Agents read state aggressively. They poll constantly, subscribe to event streams, and simulate heavily before submitting. RPC infrastructure built to serve dashboards and wallet apps falls over when a single serious agent spins up. Rate limits become the binding constraint before gas does, and the protocols that scale agent activity are the ones that treat read throughput as seriously as write throughput.
Fee design. Human fee tolerance is high because humans transact rarely. A 0.3% swap fee is invisible when you trade once a week. An agent executing 10,000 transactions a day cannot absorb percentage-based fees at that volume — the margin on any given trade is measured in basis points.
Fees need to be flat, small, and predictable, or the agent routes elsewhere. The Collisons describe a memecoin frenzy in 2025 that delayed payouts for one Stripe-Bridge user by over 12 hours and spiked per-transaction prices 35x. "Gas spikes during volatility" is a UX annoyance for humans and an economic impossibility for agents. Any design that exposes agents to unbounded fee variance during exactly the moments they need to act is broken by construction.
Matching engine and state access throughput. Agents do not queue politely. A portfolio rebalancer and an arbitrage agent both want their transactions first, and both will pay to make it happen. The market microstructure that results — auction-based inclusion, private orderflow, sophisticated MEV mitigation — has to be native to the protocol, not bolted on. Protocols that solve inclusion fairness and state-access contention at the base layer will attract orders of magnitude more agent activity than those that punt these problems to the application layer.
Deterministic execution. Agents do not handle surprise well. When a transaction simulates successfully and then reverts onchain because state moved in the intervening blocks, an agent has to reason about failure, retry logic, and expected cost. The protocols that minimize that failure surface — predictable latency, deterministic state transitions, clear ordering guarantees — are the protocols agents will prefer. Probabilistic UX was tolerable for humans. It is a tax on agents.
A protocol can compensate for one of these weaknesses with strength in another. It cannot compensate for failure on all four. The design choices that define agent-ready infrastructure compound — the same architectural decisions that make an L1 friendly to machine actors also tend to make it hostile to legacy DeFi assumptions about gas auctions, finality expectations, and block space pricing. Agent-native protocols are not just faster chains. They are different chains.
The first-mover advantage in agent infrastructure
Once agents arrive, network effects compound fast. High-volume agents concentrate where liquidity is deep and execution is cheap. Liquidity follows volume. Volume attracts more agents. The equilibrium is highly concentrated — probably more concentrated than human DeFi ever was, because agents care less about brand, UX, or community narrative, and more about the clean-room economics of execution. A difference of a few basis points per transaction, multiplied across millions of transactions per day per agent, is the entire P&L of a strategy. Agents will route accordingly.
That dynamic rewards infrastructure built for agent volume before the volume arrives. Retrofitting a blockchain for a million TPS is not a software upgrade. It requires rebuilding execution, data availability, state management, and consensus in concert. The protocols that will host agentic finance in 2030 are being architected now. The chains that wait for demand to show up will be building on top of an architecture that cannot absorb it.
This is the specific positioning Nexus is designed for. Verifiable computation at scale, paired with an execution layer built for machine actors rather than human attention, is not a speculative optimization.
It is the non-negotiable infrastructure requirement that agent adoption makes explicit. The volume we are building for is not the volume that exists today. It is the volume that will exist when agents hold wallets — and the agent wallets outnumber the human ones.
The infrastructure gap is the real story
The agent transition is often framed as a user interface change — replacing buttons with natural language, dashboards with conversations. That framing misses the actual shift. The interface change is cosmetic. The infrastructure change — from blockchains paced by human attention to blockchains paced by compute — is the change that matters.
Every major blockchain was designed for human users. AI agents transact at machine speed and need cryptographic proof of execution. That's a categorically different infrastructure problem.
Daniel McGlynn
