KOSUNI CRYPTO
Korea's Crypto Pulse, in English
$BTC $ETH $XRP $SOL $DOGE
DeFi
Neutral

Why Agentic Finance Needs an Execution Harness

Published September 29, 2026 12:36 PM · 5 views $X402
Why Agentic Finance Needs an Execution Harness

"Moving this asset to another chain, paying data fees there, and then retrieving the receipt."

This is a simple request for a person. But for an AI agent to complete it, they must choose a bridge, compare quotes, create approval transactions, calculate gas on the destination chain, recover stalled transfers, prevent double payments, and verify that the initial request was fulfilled.

Stories about the era of AI spending often begin with "just give agents wallets." For simple payments, this is not wrong. When an agent pays API fees via x402, it’s a matter of price check, approval, transfer, and receipt—done. But the request above isn’t a payment; it’s an agentic finance workflow where payment is just one step. This distinction matters: explaining all value transfers through complex financial systems makes simple payments seem harder, while calling every financial workflow a payment hides the difficulty of execution.

1. The Boundary Between Payments and Finance
Agentic payments still involve judgment—what to buy, which provider to use, how much within limits. But payments remain limited to condition checks, approvals, value transfers, and proof returns. For services like x402 sending payment conditions, agents sign and send requests with proof back through wallets or payment middlewares.

Agentic finance starts when agents take responsibility for financial decisions and outcomes—swapping, bridge path selection, approval limit management, treasury rebalancing, lending, hedging. Here, systems must preserve intent across multiple steps and prove the final state matches the initial goal. Commerce overlaps with finance but remains a separate category; even in its flow, payment is just one step. The distinction is simple: payments move value, while finance decides how to allocate, safeguard, and operate capital.

2. The 7-Layer Stack of Agentic Finance

[Image: Agentic Finance 7-Layer Stack — Layer Roles and Key Players (Source: Aomi Labs Research)]

No single model, wallet, protocol, or chain can represent all of agentic finance alone. The process from goal to settlement spans seven layers.

Layer 1 handles agent services receiving user goals and consent (ChatGPT, Jemina, Amazon Ruffes, PayPal). Layer 2 manages workflows, memory, retries, and state via runtime. Layer 3 translates plans into actual orders or transactions through domain execution (1inch, DeBridge, Uniswap, Kraken). Layer 4 covers wallets and permissions—determining who can act on behalf of whom and whether to sign (Coinbase, MetaMask, OKX, Privy, Turnkey, Safe, Fireblocks competing; Google AP2 and Visa/Mastercard setting standards). Only layer 5 deals directly with payments: x402, MPP, ACP, UCP handle price, payment method, and proof exchange. Layer 6 includes stablecoins and treasury (Circle, Tether, Stripe, bridges). Layer 7 is settlement rails (Base, Solana, Ethereum L2s, card networks, banking systems). Security, compliance, simulation, and mediation span all seven layers.

This map assigns responsibility, not companies. A company may cover multiple layers but with distinct roles per layer. Payment protocols coordinate payment steps; wallets approve or reject requests. But neither is responsible for verifying whether a multi-step financial goal was completed correctly.

3. The Boundary of Approval
While the front end of finance workflows can be flexible—interpreting goals, choosing paths, drafting proposals—the moment a request reaches spending authority, decisions must become decisive. Signers must verify unchanged requests against identity, consent, limits, and policies; settlement rails process intent without reinterpretation. IMF’s 2026 agentic payment model separating "intent coordination," "control and approval," and "settlement" follows the same logic: models freely propose, wallets independently reject, runtime prepares evidence but doesn’t assume authority.

4. Current Market Position
Narrowly defined payments are rapidly maturing into products. Wallets evolve from key vaults to programmable permission systems. Coinbase isolates keys in secure enclaves with limits; MetaMask maintains self-custody while mandating pre-signature simulation. Circle and Crossmint bundle wallet, funds, and payment protocols together. Protocols don’t converge but coexist by role: x402/MPP handle automated payments, ACP/UCP manage commerce flows, AP2 and card network protocols address permission delegation and identity.

Channelis data shows Base’s x402 transactions exceeded 100 million by Q1 2026, with Solana surpassing 35 million. However, some activity was speculative. This demonstrates technology adoption—not yet evidence that complex autonomous finance demand has taken root. Adoption starts with digital use cases where machines read prices, can be reversed via API/data calls—payments through API/data calls, not cross-chain routing or treasury management which carry higher execution risks and regulatory challenges. Stablecoins are key settlement tools but don’t solve agent identity, refunds, taxes, or transaction composition.

5. What Execution Harnesses Handle

[Image: Execution Lifecycle 7 Steps and Approval Boundary — Where Harnesses and Wallets Operate (Source: Aomi Labs Research)]

Not all payments need execution harnesses. They become necessary when multiple actions depend on each other, paths must be chosen, states change continuously, partial failures or irreversible executions occur, or receipts alone can’t prove outcomes.

Returning to the initial request: wallets set spending limits, simulators run transactions beforehand, chains prove finality. But no one among them ensures that the entire path from route selection to final balance matches the original goal. Currently, models fill this gap with reasoning—agents reread documents, recheck approval rules, retry on failure. This consumes tokens and time, risks double payments as quotes change, and introduces uncertainty.

Execution harnesses take over this role within runtime. They interpret goals, choose tools and paths within controlled scope, create precise transactions, simulate before permission requests, submit only one request for review, and compare final state against initial intent. Unlike toolkits, they own state, recovery, completion; unlike wallets, they don’t assume signing authority. The right to reject remains with the wallet until the end. Thus, harness outputs are not just successful transactions but evidence chains linking intent through approval, receipts, and final state.

6. Why Harnesses Are Needed Even With Better Models
Will smarter models make this layer obsolete? No—approval sequences, duplicate execution prevention rules, receipt formats, balance checks don’t change with each execution. These should be embedded in verified software; model reasoning should focus on actual judgment calls. Legal and medical AI have already taken this path: Harvey wraps general models in workflow, citation requirements, human review; Abridge ties draft medical records to original sources. Repetitive professional work is moving from prompts to software. Finance involves irreversible asset movement, so software must directly handle configuration simulation, duplicate prevention, recovery, and mediation.

This claim requires measurement: compare harness vs. no-harness scenarios under identical models, tasks, signing policies, market conditions—measuring tokens, time, cost, human intervention per verified outcome. Track success rates, risky proposal block rates, simulation-execution alignment, final state accuracy. Benefits must persist even in unfavorable conditions like outdated quotes or stalled bridges. Harnesses that hide failures or just increase retries don’t prove value.

7. The Next 2–3 Years
In under a year, agent spending policies and transaction security will become standard wallet features; processors will support both x402 and MPP. This is agentic payment evolution. Bigger opportunities emerge on top: agents move between services with delegated authority, routers choose stablecoins, cards, or banks, treasury and compliance must link all actions to correct agent, task, result. Surviving systems separate proposal, execution, approval, settlement while linking them via evidence. At that point, agentic payments become embedded features, and agentic finance becomes the broader operating system where payments happen.

Agentic payments create paths for agents to spend money. Agentic finance enables agents to make financial decisions within delegated authority and execute them. Better models yield better judgments. Harnesses ensure those judgments result in verifiable outcomes without usurping wallet permissions.

Aomi Labs is developing infrastructure that turns AI agent intent into on-chain verified execution across both payment and non-payment actions. Focusing on runtime (layer 2) and domain execution (layer 3), they connect rather than replace wallets, payment protocols, or settlement rails. The "with vs. without harness" comparison remains a validation method, not yet measured results.

The company will meet Korean builders at KBW events: TRUST404 Hackathon (Sept 29), EASYCON (Oct 1), Dev Korea Mixer (Oct 2).

Korean Source

This article is an English localization of a Korean-language crypto news report. Original headline: 에이전틱 금융에 '실행 하네스'가 필요한 이유