AI / Agents
The Agentic Future and the Token Economy
A founder reads three things in the same week. A frontier lab declares that answering questions will soon be the least useful thing AI can do. A research firm predicts that more than 40% of agentic AI projects will be cancelled within two years. A venture partner frames the "agent economy" as a multi-trillion-dollar opportunity. All three are describing the same moment, and all three are, in their own way, correct.
The agentic future is no longer a question of "if." It is a question of "when, and how well." The infrastructure shipped: open protocols for connecting models to tools, agent-to-agent communication standards, and a full stack of payment rails for machine-to-machine commerce. The capital arrived. The consensus formed. And yet the same year produced some of the most credible skepticism the field has seen. This article is an attempt to hold both truths at once, and to extract the parts that are actually actionable for someone building today.
TL;DR
Three claims, stated plainly.
First, the agentic thesis is mainstream consensus among labs, hyperscalers, and venture firms, but it sits in the trough of the hype cycle. Real infrastructure shipped in 2025 and 2026, yet Gartner predicts over 40% of agentic projects will be cancelled by the end of 2027, and a RAND meta-analysis of 65 enterprise AI initiatives found roughly 80% fail to deliver value, at twice the failure rate of non-AI IT projects. The future is agentic; it will arrive unevenly and later than the loudest voices imply.
Second, "tokens as currency" is directionally right but mechanically wrong. LLM inference tokens have become the dominant unit of cost, labor, and value accounting in AI. But the careful consensus, including Satya Nadella's own framing, is that inference tokens are a unit of account and labor, not a medium of exchange. The actual money flowing between agents is overwhelmingly stablecoins, not LLM tokens and not speculative crypto "AI agent" tokens.
Third, building software for agents to consume is the highest-conviction, most actionable part of the thesis. Anthropic's Model Context Protocol has become the de facto standard, with more than 10,000 active servers and 97 million monthly SDK downloads, and a full agentic-commerce payment stack emerged across 2025 and 2026. For builders, value concentrates in infrastructure, protocols, and agent-native interfaces.
Key findings
The consensus has shifted from debating whether agents matter to debating when and how well they will work. Every frontier lab, every hyperscaler, and the major venture firms now organize roadmaps around agents. The economic engine underneath them is the inference token, and its price is collapsing roughly 10x per year, a phenomenon a16z named "LLMflation." But agents consume vastly more tokens than chatbots, so total spend rises even as unit cost falls. "Tokens as currency" is real as an accounting metaphor and contested as literal money. The medium of exchange in agent economies is converging on stablecoins. MCP won the protocol war for agent-to-tool connectivity. Money is flowing to agent-native startups, heavily concentrated at the top. And the skeptics are credible enough to temper every timeline.
The rest of this article works through each of these in turn.
The agentic thesis
"Agentic AI" describes systems that plan, reason, use tools, and execute multi-step tasks with minimal human input. The useful mental model is a shift from the read path, where AI analyzes and answers, to the write path, where AI acts. This is distinct from a chatbot, which maps a single prompt to a single response, and from rule-based automation, which follows fixed scripts. Multi-agent systems extend the idea: an orchestrator coordinates specialized sub-agents.
What the leaders say is remarkably aligned. Anthropic frames multi-agent orchestration as a foundational 2026 trend. Microsoft's Satya Nadella describes Azure as an "AI factory," predicts a collapse of traditional SaaS into agent-first software, and argues that a company's future depends on encoding its tacit knowledge into its own models. NVIDIA's Jensen Huang frames "AI factories" that manufacture intelligence by turning tokens into monetizable output. The venture framing matches: a16z casts falling inference cost as the enabling condition for agents, while Andrej Karpathy, on the skeptical end of the optimists, calls this the "decade of agents," not the year, a deliberate deceleration.
The skeptics are not cranks. Gartner's June 2025 prediction is blunt: over 40% of agentic AI projects will be cancelled by the end of 2027 due to escalating costs, unclear business value, or inadequate risk controls. Gartner also estimates that only around 130 of thousands of vendors offer genuine agentic features, with the rest engaged in "agent washing," rebranding chatbots and RPA as agents. RAND's meta-analysis of 65 enterprise AI initiatives found an 80.3% failure rate, twice the rate of non-AI IT projects: 33.8% abandoned before production, 28.4% reaching production but delivering no value, 18.1% never recouping costs. A 2026 survey found only about 17% of organizations have actually deployed agents, though most plan to within two years.
The honest synthesis: the destination is widely agreed, the route is contested, and the arrival time is later than the marketing suggests.
Tokens as economic units
The strongest version of the "tokens as currency" claim has real intellectual weight behind it. An academic survey on token economics argues that tokens have become the economic primitive of agentic AI: the fundamental unit by which intelligence is produced and measured, analogous to the kilowatt-hour of the industrial age or bandwidth in the information age. The shift is real. Enterprise AI economics is moving from fixed per-seat SaaS licensing to variable, usage-based token metering, a transition often summarized as "seats to tokens." Nadella urges firms to build "token capital" alongside human capital.
flowchart TD
T["LLM inference tokens"] -->|meter the work| ACC["Unit of account and labor:<br/>how much intelligence was deployed"]
PAY["One agent pays another"] -->|settled in| EX["Stablecoins, USDC:<br/>medium of exchange"]
ACC -. priced in tokens, paid in dollars .-> EX
But the careful position is that a token is a unit of account, not a currency. The clearest formulation: tokens are not the new currency; they are the new unit of labor, the new unit of cost, and increasingly the new unit of compensation. The FinOps community calls the token "the atomic unit of AI value," but a billing unit nonetheless. A token is roughly four characters of data, not money, and the same token count can reflect wildly different underlying economics depending on the model and provider. The reconciliation worth remembering: while LLM tokens themselves are unlikely to become currency, stablecoins are emerging as the currency of the AI agent economy.
Sam Altman's actual framing is often misquoted here. His claim is about compute and intelligence, not a tradeable LLM token. He has called compute "the currency of the future" and "the most precious commodity in the world," and argues that "intelligence too cheap to meter" is within reach, citing a roughly 10x annual decline in the cost per unit of intelligence over five years.
One more distinction matters because the words collide. LLM tokens are sub-word inference units. Crypto "AI agent" tokens are speculative cryptoassets attached to agent projects. They share a name and almost nothing else. The AI-agent crypto sector reached real market caps, but most of those tokens trade well below their peaks, and real adoption is thin. Do not build a thesis on LLM tokens becoming a tradeable currency.
| Framing | What it claims | Verdict |
|---|---|---|
| Token as unit of account | Per-token billing meters AI value | Defensible, widely adopted |
| Token as unit of labor | Tokens measure deployed intelligence | Defensible, the strongest claim |
| Token as medium of exchange | Agents pay each other in LLM tokens | Not supported; stablecoins fill this role |
| LLM token as crypto asset | Inference tokens are tradeable currency | False; conflates two unrelated things |
Building software for agents to consume
This is the part of the thesis with the least hype and the most signal. If agents are going to act, they need standard ways to reach tools, data, and each other, and those standards consolidated fast.
flowchart LR
subgraph BEFORE["Point to point: M x N integrations"]
a1["Model"] --- b1["Tool"]
a1 --- b2["Data"]
a2["Model"] --- b1
a2 --- b2
end
subgraph AFTER["With MCP: M + N integrations"]
m1["Model"] --> H[("MCP")]
m2["Model"] --> H
H --> t1["Tool"]
H --> t2["Data"]
end
BEFORE ==> AFTER
The Model Context Protocol, introduced by Anthropic in November 2024, became the de facto standard for connecting models to tools and data. It is often described as "USB-C for AI": a client-server standard that collapses the M-times-N integration problem to M-plus-N. The adoption arc is decisive. OpenAI adopted it across its Agents SDK and ChatGPT in early 2025; Google DeepMind confirmed Gemini support. By the time Anthropic donated MCP to the Linux Foundation's new Agentic AI Foundation in December 2025, it had more than 97 million monthly SDK downloads, over 10,000 active servers, and first-class support across ChatGPT, Claude, Cursor, Gemini, Microsoft Copilot, and VS Code. The foundation was co-founded by Anthropic, Block, and OpenAI, with backing from AWS, Google, Microsoft, Cloudflare, and Bloomberg. Neutral governance is the point: it signals MCP is now infrastructure, not a single vendor's bet.
A complementary layer handles agent-to-agent communication. Google's Agent2Agent protocol, donated to the Linux Foundation with 50-plus partners, lets peer agents discover and delegate to one another. The clean division of labor: MCP handles the vertical connection between a model and its tools, while A2A handles the horizontal communication between peer agents.
The web itself is sprouting agent-facing standards. The proposed llms.txt convention gives agents a curated Markdown summary at a domain root. Microsoft's NLWeb turns a site's structured data into a conversational endpoint. OpenAI's AGENTS.md provides project instructions for coding agents and already appears on tens of thousands of repositories. The broad shift is from human-first interfaces, full of boilerplate HTML and rendering scripts, toward Markdown-first, API-first, schema-rich surfaces an agent can parse cheaply. Just as the last era rewarded search engine optimization, this one rewards what is now being called Agent Experience optimization: exposing endpoints, publishing structured data, and designing for machine consumption rather than human clicks.
Agent-to-agent commerce and payments
If agents transact, they need to pay. The payment stack layered rapidly across 2025 and 2026.
| Protocol | Backer | What it does |
|---|---|---|
| x402 | Coinbase | Revives HTTP 402 for instant stablecoin (USDC) payments over HTTP |
| AP2 | Authorization layer using signed "mandates" as verifiable credentials | |
| ACP | OpenAI and Stripe | Powers in-chat checkout; merchant stays merchant of record |
| UCP | Google and Shopify | Broader discovery-to-post-purchase commerce lifecycle |
| TAP | Visa | Signs agent identity into HTTP headers |
The reason crypto rails fit machine commerce is structural, not ideological. Traditional rails carry fixed fees of roughly five to fifteen cents per transaction, which makes sub-thirty-cent micro-purchases (per-inference, per-API-call) uneconomic. Stablecoins on fast layer-two networks settle in under two seconds for a fraction of a cent. Agents also cannot easily hold bank accounts or credit cards, so programmable wallets fill the gap. With stablecoin supply well above $300 billion, the liquidity exists.
The caveats here are heavy, and worth stating before anyone bets on the numbers. The headline x402 volume figures are badly inflated. An analysis by Artemis found that roughly 48% of transactions and 81% of transaction volume were linked to gamed activity, namely self-dealing and wash trading. Measurements of the same window diverge by an order of magnitude across platforms. OpenAI scaled back its in-chat checkout in early 2026 after adoption stalled at a handful of merchants. The rails are real; the traction is mostly not yet.
pie showData
title x402 transaction volume, late 2025 (Artemis estimate)
"Gamed: wash trading and self-dealing" : 81
"Legitimate" : 19
The market opportunity
Market-size estimates for agentic AI span an almost comic range, from a few billion dollars today to figures in the tens of trillions by 2030, depending entirely on definition. Treat all of them as directional forecasts, not facts. The more useful signal is where capital actually concentrates.
The funding mix is rotating. Infrastructure's share of mega-rounds fell from about 65% in 2024 to 38% in 2025, while vertical applications rose from 12% to 30% and horizontal workflow agents from 5% to 18%. The headline raises are concentrated at the top: Sierra raised a $950 million round in May 2026 at a $15.8 billion valuation, having just hit $200 million in ARR; Anysphere, maker of Cursor, raised $2.3 billion at a $29.3 billion valuation in late 2025; Decagon tripled to a $4.5 billion valuation in January 2026. A cluster of dedicated agent-payment and identity startups raised tens of millions. Across the category, early 2026 ran at roughly twice the capital and three times the deal count of early 2025, with the top ten deals capturing nearly 80% of the money.
The business-model shift underneath all of this is from per-seat pricing to outcome and usage-based pricing: Sierra charges per resolution, Decagon per conversation. Token and usage metering is becoming its own infrastructure category.
Recommendations
Staged by conviction, for a builder acting on this thesis.
The highest-conviction move, to do immediately, is to build agent-native now. Ship an MCP server for your product or data; it is the single lowest-regret action, with cross-vendor buy-in and neutral governance. Start read-only, with scoped auth, audit logs, and human-approval gates for any write action. Publish llms.txt and complete your structured data markup. Adopt usage or outcome-based pricing and instrument token metering from day one, because agent token consumption scales with enthusiasm, not with value delivered. Build cost governance (model-tier routing, prompt caching, budget caps, context pruning) into the architecture rather than bolting it on later.
The high-conviction move over the next six to eighteen months is to position for agent commerce. If you sell anything, implement the commerce protocols where your buyers are, and keep product feeds real-time and schema-complete. Experiment with x402 for machine-to-machine API monetization, but treat current volume metrics skeptically and do not bet the business on crypto-rail volume yet. Watch the agent identity and trust layer closely; the next durable infrastructure value likely concentrates there, not in the payment rail itself, which is trending toward commodity.
The speculative tier is simple: do not conflate LLM tokens with crypto "AI agent" tokens. If you engage stablecoin rails, build on USDC and facilitators rather than betting on any single chain or native token.
A few thresholds would change the recommendation. Accelerate if MCP adoption keeps compounding, if de-gamed x402 volume shows sustained retention, and if a frontier model reliably sustains multi-day autonomous tasks. Decelerate if Gartner's cancellation prediction materializes in your vertical, if frontier per-token prices rise sharply as the subsidy era ends, or if agentic-commerce checkout keeps stalling.
Caveats
The hype-versus-reality gap is large and well-documented; Gartner and RAND are credible counterweights, and most enterprise agent deployments remain narrowly scoped pilots. Payment-volume metrics are unreliable and heavily inflated by wash trading. "Tokens as currency" is mostly metaphor; the defensible claim is tokens as a unit of account and labor, and the literal machine-to-machine money is stablecoins. Market-size numbers are forecasts with enormous dispersion and lean on "could" and "projected" language. Standards are still consolidating, and some competing commerce protocols will not survive, so it is safest to build on the layers with cross-vendor neutrality and to hedge on the contested ones. Finally, source quality varies: many adoption statistics originate from vendor blogs, and a few specific figures could not be independently triangulated, so they should be read as indicative rather than precise.
The shape of the conclusion is this. The future is agentic, the token is its meter, and stablecoins are its money. The builder who treats those three sentences as distinct, rather than collapsing them into a single slogan, will make better decisions than the market around them.
References and good reads
Primary sources and protocols
- Anthropic. Donating the Model Context Protocol and establishing the Agentic AI Foundation.
- Model Context Protocol. MCP joins the Agentic AI Foundation.
- a16z. Welcome to LLMflation: LLM inference cost is going down fast.
- Epoch AI. LLM inference prices have fallen rapidly but unequally across tasks.
- Coinbase. x402 developer documentation.
Skeptical and critical takes
- Gartner. Over 40% of Agentic AI Projects Will Be Canceled by End of 2027.
- RAND analysis, as summarized in Why most enterprise AI projects fail.
- CoinDesk. Coinbase-backed AI payments protocol wants to fix micropayments but demand is just not there yet.
Market and funding