Why crypto rails won the machine economy
Three things that banks keep separate collapse into one object on-chain:
- Identity — a key, a smart account, or an ERC-8004 passport NFT, created in milliseconds with no protocol-layer KYC.
- Money — overwhelmingly USDC: ~98.6–99% of settlement on the agent rails that can be measured.
- Programmable rules — spend caps, allowlists, session keys, EIP-7702 delegation, vaults, escrow. Policy enforced in code, not in a prompt.
TradFi cards still win on refunds, chargebacks, and large consumer purchases. Stablecoins win on sub-dollar, high-frequency, cross-border, machine-to-machine settlement. That split is already visible in the data: across $73M and 176 million agent payments measured from May 2025 to April 2026, the average ticket was ~$0.31 and 76% of payments sat below the $0.30 card-fee floor (Keyrock, with Coinbase and Tempo data). Cards cannot price a three-cent API call. That is why machine commerce chose stables →
The five behaviors that exist today
Ranked by receipt count, not by press release:
01
Paying for tools
An agent requests a resource, the server answers HTTP 402 Payment Required, the agent signs a USDC authorization, a facilitator settles, the server delivers. Tens to hundreds of millions of receipts. Tickets in cents.
02
Holding & moving stables
Agent wallets with scoped spend from Coinbase, Privy, Binance, Circle, Trust Wallet, Turnkey, Cloudflare, MoonPay, Safe. The common funding pattern: USDC in, USDC out, gas sponsored so the wallet never holds a gas token.
03
DeFi copilots
Natural-language swap, rebalance, limit, perps. Mixed autonomy — humans still approve size. Bankr routes >$300M weekly Uniswap volume on Base as an execution surface.
04
Yield on idle USDC
Treasury agents refuse to sit on 0% cash between jobs — DeFi yield routers (ARMA, ZyFAI, Yield Seeker) and regulated ERC-4626 RWA vaults like the IXS Agentic Vault, where deposits inside policy need no human signature.
05
Agent-to-agent jobs
Escrowed work under draft ERC-8183: Open → Funded → Submitted → Completed | Rejected | Expired. Early, but this is how agent GDP becomes more than API micropayments.
⚠
Getting attacked
A documented 2026 case: an attacker manipulated an agent into transferring ~$175k via hidden instructions — not a stolen key, not a broken AMM. Incidents are a first-class category of agentic finance, not an afterthought.
The stack, as one graph of layers
Intent (LLM / rules / mandate)
→ Identity (wallet + ERC-8004 + optional KYC/TEE)
→ Policy (caps, allowlists, session keys, EIP-7702)
→ Orchestration (MCP / A2A / skills)
→ Execution (DEX, CEX API, vault, bridge, x402)
→ Settlement (USDC / CCTP / account abstraction)
→ Evidence (tx, reputation, validation registry)
If a product only covers one layer, it is a feed, not the graph. Each standard is an edge type — REGISTERED_AS, DELEGATES_CODE_TO, PAYS, DEPOSITS, REVIEWS, VALIDATES. The graph model in full →
What agents are not doing at scale yet
- Fully unconstrained “maximize yield across every chain” intent execution.
- Comparable on-chain reputation that predicts performance — 8004 scores are signals, not rankings.
- Multi-stablecoin settlement. It is a USDC monoculture, which is both convenience and systemic risk.
- Validation-registry usage — some 30-day explorer windows show single-digit validation requests.
- Clean liability assignment when an agent is prompt-injected.
Honest sizing: measured rails settle tens of millions of dollars; labeled AgentFi TVL is low-single-digit millions per chain against a $311.1B stablecoin float. The gap between those two numbers is the market — and the thing our panels track.