The GDP Unlock: Agents and the Next Exponential

From cheaper inference to useful work

The input cost of machine cognition is collapsing at a rate with no industrial precedent. Epoch AI's price tracking finds that the price of a fixed level of LLM capability has fallen between 9x and 900x per year depending on the benchmark, with a median around 50x per year. Stanford's 2025 AI Index measured the cost of GPT-3.5-level inference falling roughly 280-fold in eighteen months. Even the conservative estimate — a16z's "LLMflation" figure of about 10x per year — would be, for any physical input, the fastest cost decline ever recorded.

Yet it is easy to confuse more output with more useful work. A model can produce a million words that nobody needs. It can also produce ten lines of code that save someone a week. The number of tokens tells us almost nothing about the difference.

An agent turns compute, software, tools, and context into an attempted result. Cheaper inference makes more attempts affordable. It does not make them good. That still depends on the brief, the tools, the security boundaries, and whether anyone can tell if the result worked.

This is why markets matter. A funded brief says that someone wants a particular outcome badly enough to pay for it. Agents supply attempts. Tests, evidence, or judgment separate the useful attempt from the rest. Payment closes the loop. Without those pieces, abundant model output may remain just that: output.

The rails exist. The dollars are still small.

The strongest evidence for this thesis is an uncomfortable pair of numbers. Chainalysis counted over 100 million x402 transactions by the first quarter of 2026, growing from near zero in mid-2025 — machine-to-machine payments are demonstrably real. But CoinDesk's reporting put the adjusted dollar volume across those transactions at roughly $15 million as of March 2026. A hundred million transactions; fifteen million dollars.

The same pattern holds one level up. Raw stablecoin transfer volume reached the trillions, but BCG and Allium's adjusted analysis puts real-economy stablecoin payments at $350–550 billion in 2025 — a fraction of the headline, and mostly B2B settlement. Transaction counts and raw throughput systematically overstate the economics.

Read one way, this is deflating. Read correctly, it is precisely the pre-unlock shape: the pipes are installed and pressurized before anything valuable flows through them. TCP/IP existed in 1983 and HTTP in 1991, but e-commerce only compounded after payments, trust, and logistics arrived — and then it went from zero to 16.4% of US retail sales ($1.23 trillion in 2025). The bottleneck was never the transport protocol. It was the coordination stack around it.

The architecture of agentic commerce

For an agent to take part in commerce, it has to do more than call an API. It must be able to find an offer, understand the terms, know whether it is allowed to pay, and receive the thing it bought. That requires five fairly unglamorous layers:

1. Discovery: the agent needs a structured description of the task or service.
2. Authorization: a wallet or mandate has to limit what the agent may spend.
3. Execution: the service needs a stable interface the agent can call.
4. Verification: the buyer needs evidence that the promised outcome arrived.
5. Settlement: payment and a durable receipt have to follow the accepted result.

The x402 documentation describes one way to handle the payment part. A server answers an unpaid request with HTTP 402 Payment Required and terms a program can read. The client signs a payment payload and tries again. If the payment checks out, the server returns the resource and settlement information. This works neatly for pay-per-request services. It does not tell you whether a larger job was done well.

Task markets begin where that neat exchange stops. Taskmarket lets a requester fund an outcome and set the rules for claiming or competing for the work. The market records bids, submissions, and acceptance before it releases payment. The protocol prices access to a capability. The market coordinates the delivery of an outcome. These are related jobs, but they are not the same job.

What history says coordination is worth

The claim that boring coordination infrastructure unlocks disproportionate value is not a metaphor; it has been measured twice.

The shipping container was trivial technology — a steel box. But Bernhofen, El-Sahli, and Kneller estimate that containerization raised trade among industrialized countries by a cumulative ~700% over twenty years, a larger effect than all tariff reductions of the era combined. The box was cheap. The standardization of ports, customs, insurance, and handling around the box was the unlock.

The internet repeated the pattern. McKinsey's Internet Matters study found the internet accounting for 3.4% of GDP across the economies studied and 21% of GDP growth in mature economies over the preceding five years — with 75% of that value accruing to traditional firms, not technology companies. The value of a general-purpose network shows up mostly outside the network's builders, and only after the coordination layers mature.

If agents follow the same curve, the interesting question is not model capability. It is how fast demand becomes legible to software, and how fast payment, verification, and recourse become routine.

Context, memory, and specialization

A general model is only one part of an agent. The agent also needs instructions, tools, state, and knowledge about the domain it is working in. The open-source Daydreams framework organizes that state into composable contexts. Lucid Agents gives TypeScript developers a way to expose an agent's capabilities as paid services.

Specialization matters because useful work happens inside constraints. A coding agent needs the repository and its test results. A market agent needs current prices, risk limits, and narrow transaction permissions. A research agent needs rules about sources and citations. The more context an agent keeps, the more continuous its work can become. But memory creates its own problems: private data, stale facts, and old mistakes that persist because nobody corrected them.

Evidence today, thesis tomorrow

The measured evidence on agent labor is genuinely mixed, and the thesis is stronger for admitting it.

On the encouraging side: the landmark customer-support study by Brynjolfsson, Li, and Raymond found AI assistance raised issues resolved per hour by 14–15% on average and 34% for novice workers. GitHub Copilot's controlled trial measured 56% faster task completion. OpenAI's task-exposure analysis, published in Science, estimates about 80% of US workers have at least a tenth of their tasks exposed to LLMs. And Anthropic's Economic Index finds enterprise API usage already running around 75% automation rather than augmentation — delegation, not autocomplete.

On the sobering side: a METR randomized trial in 2025 found experienced open-source developers were actually 19% slower with AI tools while believing they were 20% faster. Micro-level task gains coexist with stubbornly flat aggregate productivity statistics — the classic J-curve of a general-purpose technology before its coordination infrastructure matures.

These facts do not prove that agents will replace broad categories of labor. They do not prove every task belongs in an agent market. And they certainly do not make token throughput a substitute for GDP — the x402 numbers above are the cleanest demonstration that machine transaction counts and economic value are different quantities. Progress should be measured by accepted outcomes, cost, speed, reliability, and how much human supervision the system still needs.

The exponential ahead

Technical capabilities become economically important when institutions help people allocate resources and keep agreements. The same is likely to be true of agents. Better models matter. So do markets, payments, evaluation, identity, and some form of recourse when things go wrong.

So the Daydreams thesis is deliberately practical. Make demand legible to agents. Make agent capabilities legible to buyers. Attach payment to a request. Keep enough evidence to decide what happened afterward. Inference will keep getting cheaper on its own — that curve needs no help from us. Whether the resulting abundance of attempted work reaches the economy depends on the mundane coordination systems this essay is about, and that is the part still up for grabs.