Daiwa Runs AI Agents Through On-Chain Trading as MAS Drafts KYA
Japan’s first securities group has run a full on-chain trading loop with AI agents, from natural-language instruction to human pre-approval and an auditable trail. In the same week, Singapore’s MAS unveiled a Know-Your-Agent protocol and Token Terminal clocked an 498.8% jump in agent-initiated USDC transfers, pushing machine-agent finance into the identity-and-authorisation standardisation phase.
On November 20, Daiwa Securities Group, Daiwa Securities and Daiwa Institute of Research jointly announced that they had run a full on-chain trading flow with AI agents in an internal test environment: multiple agents receiving natural-language instructions collaborated to assemble candidates from several decentralised exchanges (DEXs), and after the user selected a route and gave prior approval, the flow of exchanging test stablecoins for test security tokens (STs) was confirmed. The core components run on open standards that do not lock in any single vendor, A2A (Agent2Agent Protocol) for agent-to-agent communication and AP2 (Agent Payments Protocol) for agents transacting and paying on behalf of users.
Read against the rest of the week, the news is not isolated: on November 19, MAS Managing Director Chia Der Jiun laid out a Know-Your-Agent protocol at the Singapore FinTech Festival, and Token Terminal data shows AI-agent USDC transfers jumped 498.8% to 11.1 million in the 30 days to November 19. Put together, the three point to one shift, machine agents are moving from “being able to spend” to “institutions being willing to let them spend”, and what decides the step is no longer the wallet, but the standardisation of identity, authorisation and audit.
1. Daiwa’s PoC: The Securities Industry’s Agent-Ordering Loop, End to End
The experiment ran in an internal test environment: a user issued a natural-language instruction, multiple AI agents collaborated to present candidates from several DEXs, and only after the user selected a route and pre-approved did the agents execute the on-chain swap of test stablecoins for test security tokens. The division of labour among the three companies: Daiwa Securities Group drives digital and AI strategy and examines future business concepts; Daiwa Securities plans agent use cases in future securities services and organises approval procedures, governance, operations and legal and compliance issues; Daiwa Institute of Research designed the overall demonstration system, built the test environment, and implemented the agent functions plus behavioural-trail and audit-log management.
The key increment this time is traceability. The three companies confirmed that the content of instructions to the agents, the content of approvals and the transaction results could be linked and recorded, letting the whole process be reviewed after the fact. In regulatory terms this maps to explainability: when a machine agent makes an ordering decision on behalf of a client, someone must be able to reconstruct why it bought, who approved, and what filled. The experiment also verified the feasibility of transactions spanning multiple DEXs, and of mechanisms that control what content and conditions agents may execute.
The next steps are deliberately incremental: under laws and regulations, internal rules and risk-management and governance requirements, the companies will examine step by step the scope of operations that can be delegated to AI agents, how humans stay involved, record-keeping and audit mechanisms, collaboration with external systems and external agents, and where responsibility lies for instructions and execution. No product or service has a timeline.
2. MAS’s KYA Protocol: A Financial ID for AI Agents
MAS Managing Director Chia Der Jiun announced at the Singapore FinTech Festival on November 19 that the regulator will develop a common “Know Your Agent” protocol. Financial institutions would use it to learn who an agent is, whom it represents, what permissions it holds, and what assurance exists about its provenance and reliability, then decide which services the agent may use, under what conditions, and open a dedicated secure access channel.
The KYA protocol builds on MAS’s Safeguards for Agentic Finance at Runtime (SAFR) framework. SAFR’s first phase produced four lessons: agents need bounded autonomy; identity and authority must be established; mandates must specify permissible actions, limits and conditions; and controls should strengthen as an agent’s discretion and potential impact increase. MAS also flagged a live risk: some users have handed financial-account login credentials to AI agents, and without sufficient safeguards institutions cannot confirm whether the agent was fully authorised or whether a transaction exceeded its mandate.
Supporting moves came in the same week: MAS repositioned the SGD Testnet as the SG Digital Ledger, an operational settlement utility, and announced that S$100 million of tokenised MAS Bills issued on November 6 had been settled in wholesale central bank digital currency, with demand more than twice the amount on offer. MAS and the Future of Finance Institute set up an Agentic Money Sandbox whose first use cases cover agent-initiated payments and corporate treasury management, where a corporate can authorise an agent to monitor liquidity and move funds with tokenised deposits.
3. The Machine Agent’s Wallet: USDC Transfers Up Nearly 6x in 30 Days
Token Terminal data shows AI-agent-initiated USDC transfers reached 11.1 million in the 30 days to November 19, a 498.8% jump from the prior window. 92% of the activity ran on three networks: Base with 4.8 million transfers, Polygon with 2.9 million, and Circle’s Arc L1 with 2.5 million. Circle’s own number is bigger: agents completed more than 52.5 million USDC transactions in the 90 days to November 9, 99.99% of the agentic transaction volume it tracks.
Transfer values typically range from fractions of a cent to a few dollars, and much of the flow runs through x402, a protocol that tucks a payment request inside an ordinary HTTP response so an agent settles an API call or a data purchase on the fly. Card networks and bank transfers were built for humans and cannot carry this kind of micropayment volume; on-chain stablecoin settlement changes the cost structure. Circle launched Agent Stack in May and its Arc L1 mainnet in September, giving agents programmable spending power and gas-free transactions.
The striking mismatch is volume versus value: transaction counts exploding while total dollar value stays tiny is exactly what you would expect from an agent economy moving from experiments to real infrastructure calls, agents buying APIs, data and inference, each payment too small for a human to bother with, but the frequency already rewriting the traffic structure of payment rails.
4. Trend Analysis: Three Events, One Picture
5. Investment View
The crypto market remains under pressure: Bitcoin fell 2.7% to about $61,500 on November 20, total market cap shrank 1% to $2.21 trillion, and the Fear and Greed Index at 12 stays in extreme-fear territory. Against the price slump, the infrastructure layer is active: USDC market cap stands at $280.7 billion with $68.8 billion in volume, and agent payments are becoming a new “non-trading use” narrative for stablecoins. Circle’s Agent Stack and Arc L1 capturing the largest share of agentic flow means its first-mover advantage in machine payments is already visible in on-chain data.
For listed companies, Daiwa’s PoC is a signal worth watching: Japanese brokers are starting to build technology stacks and governance frameworks for the agent era, and securities service design and customer touchpoints could be reshaped. But as Daiwa itself admits, without sorting out authorisation scope, liability boundaries and compliance, scaled deployment will not arrive. Investment cues at this stage sit more in the “infrastructure plus identity-authorisation” layer than in the monetisation of specific transaction volume.
6. Conclusion: The Machine Agent’s Financial Entry Ticket Is Auditable Authorisation
This week’s three events give the same answer: whether an AI agent can truly enter the financial system depends not on how smart the model is or how convenient the wallet is, but on verifiable identity, bounded authorisation and auditable behaviour. Daiwa provides the engineering sample, MAS provides the regulatory framework, and the USDC data provides the proof of traffic. For practitioners, two directions are worth tracking, the convergence progress of identity-and-authorisation standards such as KYA, A2A and AP2, and the timetable on which licensed institutions, brokers, banks and payment providers push agent access channels into production. Where the two lines meet will define the point at which machine-agent finance crosses from pilots to scale.
- BigGo Finance — Daiwa Securities Group Tests AI Agents Trading Digital Securities via DEX
- Unite.AI — MAS Sets Out Know-Your-Agent Protocol and SG Digital Ledger Plans
- MAS Media Release — National AI Missions for Financial Services
- Lianhe Zaobao — AI agents can move funds faster; MAS watches the risk
- KuCoin — AI Agent USDC Transfers Surge 498.8% to 11.1 Million in 30 Days
- Crypto Cortex — AI agent USDC transfers jump 498.8%
- CoinGabbar — Crypto News Nov 20: Bitcoin Drops Amid Japan Risk
- Web Pulse — Singapore’s MAS upgrades tokenized settlement rails

发表回复