What happens when AI agents become economic actors?
Intelligence can now act. Money can now move. This project investigates the convergence: the evidence, the scenarios, and the intellectual honesty about what we don't yet know.
Read Research 001: What Money Will AI Agents Use? →For most of economic history, every financial transaction required a human decision.
AI systems are gaining agency: the ability to take actions, not just generate information. Money is becoming programmable; machine-readable, digitally native, capable of moving through software without human initiation.
These two trajectories are converging.
The research question
Historically, software could recommend, calculate, optimize and automate. These are powerful capabilities, but they operate within decisions made by humans.
The relevant research question begins when software can increasingly:
- Initiate transactions without human instruction
- Choose counterparties and evaluate terms
- Pay for services and resources
- Hold and manage monetary assets
- Allocate capital across competing uses
- Hire other agents for specialized work
- Borrow against future revenue or collateral
- Invest in assets, infrastructure, or other agents
- Coordinate economic activity with other autonomous agents
Each of these represents a shift from tool to economic actor. Whether this shift is happening at meaningful scale, and what infrastructure it requires, is what this project investigates.
Research 001
What Money Will AI Agents Use?
The first investigation asks what monetary instruments are best suited for autonomous agent transactions and reserves. The findings reveal a significant gap between infrastructure investment and observed autonomous activity.
The infrastructure is being built
- Visa, Mastercard, Stripe, and PayPal have all shipped agent payment products
- The x402 protocol reports 205 million agent payment events
- USDC accounts for 98.6% of settlement value among measured crypto-rail agent payments
- Circle's Agent Stack supports nanopayments down to $0.000001
Sources: Visa TAP, Mastercard Agent Pay, Stripe MPP, x402 protocol metrics, Keyrock (176M transactions, May 2025 – Apr 2026)
The counter-evidence
- An estimated 0.6–7.5% of x402 settlements represent genuinely autonomous AI activity
- Approximately 48% of x402 transaction volume is attributable to wash trading
- TRM Labs estimates genuine autonomous agent-to-agent commerce at $5,000–$11,000 per month globally on crypto rails
- The vast majority of AI spending flows through conventional billing (credit cards, invoices, cloud subscriptions), with a human paying the bill
Sources: TRM Labs filtered x402 data (removed self-payments, bulk flows, sellers with <10 distinct buyers), Artemis wash-trading analysis (2026)
The gap between infrastructure investment and identified autonomous commerce is the central finding. Significant corporate investment is flowing into agent payment infrastructure. Yet the observable evidence for genuine autonomous commerce is considerably more limited, and the economic significance remains uncertain.
Read the full investigation →Defining economic agency
Not all "agent payments" are the same thing. This investigation distinguishes three levels of AI economic activity; they are almost always conflated.
| Level | Description | Who initiates | Who pays | Human approval | Current scale |
|---|---|---|---|---|---|
| 1. Operator-billed consumption | A company pays its AI API bill through conventional billing | Human operator | Human operator | Implicit (subscription) | ~$70B+ ARR |
| 2. Delegated payment | An agent uses a corporate card within human-set spending limits | Agent, within bounds | Agent, human-liable | Pre-authorized limits | Emerging |
| 3. Autonomous commerce | An agent selects counterparties, negotiates terms, and settles independently | Agent | Agent | None | $5K–$11K/mo |
The vast majority of what is called "AI agent payments" in 2026 is Level 1: conventional software procurement with an AI model in the loop. The research questions that define this project (identity, authorization, monetary instruments, treasury behaviour) emerge primarily at Level 3.
Transaction money vs. reserve money
If an autonomous agent earns revenue, it faces the same fundamental division every economic actor faces: the money used for immediate payment may serve different purposes than money held as reserves.
This distinction matters because the properties that make a monetary instrument good for transactions (acceptance, low cost, fast settlement) may differ from those that make it suitable for reserves: value preservation, low volatility, limited counterparty exposure.
Transaction needs
- Acceptance
- Must be accepted by counterparties
- Settlement speed
- Fast or instant settlement preferred
- Transaction cost
- Must be viable for small transactions
- Programmability
- Machine-readable, API-accessible
Current candidates: conventional billing, stablecoins (USDC), card-network agent APIs, Lightning
Reserve considerations
- Value preservation
- Resistant to purchasing-power erosion
- Volatility
- Predictable value over holding period
- Counterparty exposure
- Limited dependence on third parties
- Liquidity
- Convertible when needed
At $5,000–$11,000/month of identified autonomous commerce, the question may be premature. But it is the question that will define the machine economy, if one emerges.
This is a hypothesis to investigate, not a conclusion. The research does not presuppose that agents will develop reserve preferences, or that any specific instrument will serve either role.
Ten investigations
Each investigation is a standing research question; findings are versioned and updated as evidence changes.
- 001What Money Will AI Agents Use?What monetary instruments are best suited for autonomous agent transactions and reserves?60 evidence records, 51 sources, 6 claims
- 002The Agent WalletWhat financial control architecture is required when software becomes an economic actor?20 evidence records, 22 sources, 8 claims
- 003Know Your AgentHow do you establish identity, trust, and authorization for economic software?25 evidence records, 20+ sources, 7 claims
- 004Can an AI Agent Own Bitcoin?What does ownership mean when software can exercise effective control over scarce digital property?23 evidence records, 39 sources, 6 claims
- 005The Autonomous CorporationCan a corporation become an autonomous economic actor?30 evidence records, 28 sources, 8 claims
- 006Stablecoins vs. LightningWhich transaction rails are best suited for autonomous agent commerce?
- 007The Agent TreasuryHow might autonomous agents manage capital?
- 008When Agents Hire AgentsWhat happens when agents become both buyers and sellers of services?
- 009Credit Without HumansCan autonomous agents become creditworthy economic entities?
- 010Machine Capital MarketsWhat happens when agents become capital allocators?
Four competing futures
We don't predict which future arrives. We map the conditions under which different futures become likely.
These scenarios are analytical tools, not predictions. Each assumes a different configuration of the economic system described above. Their evidence relationships reflect Research 001 findings and will evolve as investigations continue.
| Scenario | Core assumption | Becomes likely if | Research 001 evidence |
|---|---|---|---|
| A: Banked Agents | Existing financial institutions successfully adapt | If major banks launch agent account products with competitive pricing by mid-2027, this scenario becomes more likely. | MixedSupported by conventional billing dominance. Challenged by the card-fee micropayment floor. |
| B: The Stablecoin Internet | Stablecoins become the primary transaction layer for machine commerce | If stablecoin agent transaction volume exceeds $1B/month by 2028 AND regulatory frameworks are favorable, this scenario becomes dominant. | MixedSupported by USDC dominance within crypto rails. Challenged by regulatory uncertainty and metrics reliability. |
| C: The Satoshi Economy | Agent-to-agent commerce becomes economically significant, and Bitcoin develops a meaningful monetary role | If agents begin autonomously holding Bitcoin reserves AND Lightning transaction volume from agents exceeds $100M/month. | ChallengingNo supporting evidence identified in sources searched. On-chain metrics reliability challenges assumptions. |
| D: The Non-Event (null hypothesis) | Agent economic activity remains trivially small or is fully absorbed by existing infrastructure | If agent transaction volume does not exceed $10B/year by 2029, this scenario becomes the most likely. | ConsistentSmall scale of identified autonomous commerce and measurement problems are consistent with this scenario. |
Evidence relationships are directional, not conclusive. They reflect one investigation out of ten. Scenario D is the formal null hypothesis: the possibility that autonomous agent economic activity remains trivially small or fully absorbed by existing infrastructure. It receives the same analytical weight as scenarios A–C.
What would change our mind
We would revise our assessment of conventional billing dominance if credible evidence emerged that non-conventional payment rails had become a material component of AI-service settlement.
We would reconsider the scale of autonomous agent commerce if independent analysis, applying comparable filtering methodology, produced estimates materially larger than the currently identified range.
We would reassess the micropayment gap if conventional payment networks demonstrated fee structures that made very small autonomous transactions economically viable at scale.
We would reconsider on-chain measurement reliability if a broadly adopted attestation mechanism demonstrated the ability to distinguish autonomous AI-agent activity from other forms of automation.
The thesis is a question, not an answer.
If the evidence shows that agent economic activity remains trivially small or fully absorbed by existing infrastructure, we will publish that finding prominently. The project must be willing to prove itself wrong.
Begin with Research 001: What Money Will AI Agents Use? →