The agent economy: infrastructure, metrics, and market tracking

Research process: this report was compiled with AI-assisted deep research. Every figure below was verified against the primary source cited before publication; claims we could not verify were cut or rewritten as qualitative statements. On-chain figures are cited to our own live instrument, which updates daily, rather than frozen in this text.

The question

For an autonomous software agent to become a meaningful economic actor, it must discover services, agree on terms, execute payment, and confirm that what it paid for actually arrived, all without a human in the loop. The Spoolis thesis is that as agentic activity scales, the last step, fulfillment verification, becomes a distinct infrastructure layer. Agents cannot safely spend real money at volume if they cannot independently verify that the purchased outcome was delivered to specification.

This report examines the evidence for and against that thesis: what agents pay for today, how agent-native payment protocols work, whether verification is separating from applications, and what the market could become. It is written to be useful to someone with no interest in Spoolis. Where the thesis could be wrong, we say so.

1. What agents actually pay for today

The clearest window into agentic commerce is the x402 protocol, an open payment standard from Coinbase built on the HTTP 402 status code. Its cumulative transaction count and observed volume are public on-chain data; we track both daily on the thesis page, with the query and coverage disclosed on the chart.

The raw numbers need careful reading. The transaction series is dominated by automated sub-cent loops: bursts of near-zero-value transfers that inflate counts without representing diverse commerce. Averages are especially misleading here, because a single day of dust transfers can drag the mean per-transaction value to a fraction of a cent. This is why our instrument reports a value-bearing transaction series and a median transaction value rather than headline counts and means, and why this report cites those series instead of quoting a number that will be stale next week.

The Machine Payments Protocol, co-authored by Stripe and Tempo, shows a different footprint: fewer transactions, skewed toward utility API consumption such as search, scraping, data enrichment, LLM routing, and CAPTCHA solving, indexed publicly by the MPPscan explorer. MPP publishes cumulative totals rather than a time series; we mirror those totals on the thesis page with that limitation disclosed.

The honest summary: machine-to-machine payment is real, measurable, and growing, and most of today's activity is still low-value automation. The interesting question is whether the mix shifts toward purchases whose fulfillment must be judged. That mix is exactly what our spend-composition series measures.

2. How agent-native payments work

Traditional payment infrastructure assumes a human: identity checks, pre-provisioned accounts, card details. Agent-native protocols remove those assumptions.

Under x402, a server answers a request with HTTP 402 and the payment requirements: price, network, address. The agent signs payment and retries the request with cryptographic proof attached. Facilitators verify signatures and settle on-chain, usually in USDC on networks with sub-second finality and negligible fees, often covering gas so the agent never touches blockchain mechanics.

MPP extends the same 402 pattern beyond crypto-native rails: agents can pay via fiat through Stripe payment tokens or via on-chain deposits on the Tempo network, which matters for enterprises that need traditional rails. And Google's Agent Payments Protocol (AP2) supplies the authorization layer above execution, with x402 explicitly integrated as its crypto extension. The industry is converging on HTTP 402 for execution while layering authorization and governance separately.

The consequence: paying is becoming the easy part. A payment moves money. It does not tell you whether the dataset was complete, the extraction was accurate, or the work met its criteria.

3. Verification is separating into its own layer

The Ethereum Attestation Service has become public infrastructure for signed, verifiable claims: a schema registry defines the structure of a claim, and attestations against those schemas can be made, verified, and revoked on-chain or off-chain. We track attestation activity on the thesis page as one signal of verification adoption.

Around it, a privacy-preserving proof stack is forming: selective-disclosure credentials and zero-knowledge proofs let a party prove a specific fact, a balance above a threshold, a completed check, without revealing the underlying data. Proof generation is fast enough for interactive use; proof verification is moving to low-cost networks and off-chain execution because mainnet Ethereum verification remains expensive.

The direction is consistent: verification is decoupling from individual applications and becoming a primitive other systems call. That is the structural claim of the Spoolis thesis, and it would be visible in the data as verification activity growing faster than payment activity. Today it is not yet established; the thesis page reports the actual state rather than assuming the conclusion.

4. Portable standards are beating bespoke integrations

Agentic commerce cannot scale if every merchant builds a custom paywall and every platform a custom acceptance flow. The pattern across the stack is convergence on portable standards: the Model Context Protocol for tool discovery and context exchange, machine-readable discovery documents that publish price and payment requirements before a request is made, public indexes like MPPscan and the x402 ecosystem's discovery layer that make services findable by machines, and IETF and W3C credential standards for proving identity and authorization across trust domains.

The same logic applies one layer up. Acceptance criteria, evidence, and verdicts are portable when they follow a standard record; they are trapped when each marketplace rebuilds them internally. That argument is developed in who decides what counted.

5. What the market could become

Institutional forecasts for agentic commerce diverge by two orders of magnitude, mostly because they define the market differently: retail agentic shopping at the narrow end, broad B2B procurement and financial automation at the wide end. We show the disagreement rather than averaging it.

Early adoption is already visible: Morgan Stanley's AlphaWise survey found roughly 23% of Americans bought something via AI in the past month, led by groceries and consumer packaged goods. And Juniper, forecasting the steepest near-term growth, states plainly that "trust will remain the number one barrier" to deployment. Every forecast above assumes agents transacting at scale; none of them work if agents cannot trust what they buy.

6. The economics of machine-scale verification

Verification is only viable if it costs meaningfully less than the transaction it protects. That threshold has been crossed for digital outcomes.

Frontier-adjacent inference has collapsed in price: DeepSeek V4 Flash lists $0.14 per million input tokens on a cache miss, $0.0028 per million on a cache hit (a 98% discount for repeated prefixes), and $0.28 per million output tokens. At those rates, running a substantial verification prompt over a large document costs fractions of a cent. Premium models cost more per token and remain the right tool where judgment quality dominates, which is why verification routing, cheap models for mechanical checks, stronger models for judgment, is an economic decision, not just a technical one.

Settlement follows the same curve: proof verification and state updates that are expensive on Ethereum mainnet cost orders of magnitude less on L2 networks and high-throughput chains, and off-chain verification on serverless edge infrastructure costs fractions of a cent per check. The combined trend means the cost to verify a digital outcome is approaching zero. Our thesis page computes this live as verification overhead: inference floor plus settlement cost, divided by the median transaction value.

7. How we track whether this thesis is right

This report originally proposed a public market-intelligence dashboard with strict rules: defined metrics, disclosed sources and coverage, medians over means, no vanity metrics, competing protocols tracked neutrally, and thresholds stated in advance so the reader can see the thesis weaken as easily as strengthen. We built it. The thesis page tracks agent economic activity, what agents buy, verification adoption, verification economics, and portable standards, each with source, freshness, and method disclosed on the chart.

The conditions we watch, stated as falsifiable directions: transaction values must rise from sub-cent automation toward amounts worth verifying; the purchase mix must shift toward outcome work whose fulfillment must be judged; verification activity must grow relative to payment activity; verification overhead must stay below a few percent of transaction value; and open standards must keep beating closed integrations.

What would change our mind

The thesis fails in identifiable ways. If merchants successfully capture agent traffic inside closed, proprietary platforms, the marketplace referee model wins again and a neutral verification layer has no room. If agentic purchasing stays permanently in sub-cent, self-evident transactions, verification never becomes worth paying for. If payment protocols absorb fulfillment verification natively, the layer exists but is not independent. We track the signals that would reveal each of these, and the thesis page is written to report them honestly if they appear.

Conclusion

Three shifts have to compound for autonomous agents to become meaningful economic actors: frictionless machine payment, which x402 and MPP are solving; affordable machine reasoning, which cached inference has delivered; and trustworthy outcome verification, which is the missing layer this report exists to watch. The payments are arriving first. The verification question, who decides what counted, is still open. That is the bet, and the instrument for watching it is live.