Infrastructure for Autonomous Agents
AI agents can only operate on what exists. The ceiling we accept becomes their ceiling.
Here’s the thesis everyone agrees on: AI agents will handle most economic activity. Portfolio management. Trading. Compliance. Billing. Logistics. Not in some distant future, 44% of finance teams already have agentic AI in production as of Q1 2026, up from 7% in January 2025. That’s a 600% increase in twelve months. Lemonade’s AI agent Jim handles 96% of insurance claims without a human. In a 14-week beta ending January 2026, over 1,000 participants created 9,500 agents that executed 187,000 autonomous crypto transactions.
The capability layer is arriving fast. The question nobody is asking loudly enough is: what are these agents actually operating on?
Because agents don’t invent infrastructure. They inherit ours. And ours is incomplete.
The Inheritance Problem
An AI agent can trade any stock on any exchange. It can execute options strategies, manage a bond portfolio, rebalance a crypto wallet, and settle cross-border payments on a stablecoin rail. Google Cloud, AWS, and Anthropic adopted the x402 protocol in October 2025, a decentralised payment standard built for machine-to-machine transactions. In March 2026, Alchemy launched a flow where an agent uses its own wallet as identity and payment source, tops up with USDC, and transacts without any human input.
The plumbing for agents to move money is materialising. But there’s a deeper problem that no amount of payment rails can solve: agents can only trade what’s tradeable. They can only see what’s visible. They can only hedge what has a market.
And most of the economy doesn’t have a market.
The Missing Markets
Think about what drives asset prices. Inflation. Employment. Consumer spending. Energy demand. Housing starts. Credit conditions. Industrial production.
Now ask: can you trade any of those directly?
Mostly, no. You can trade the effects, equities, bonds, commodities, but not the causes. An agent can buy S&P 500 futures, but it can’t buy “CPI will come in above expectations.” It can trade Treasury bonds as a proxy for interest rate expectations, but it can’t directly trade the employment report or GDP growth.
This is starting to change. CME Group now offers event contracts on economic indicators including CPI and GDP across all 50 states, alongside their traditional futures benchmarks. Kalshi, the CFTC-regulated prediction market, provides tradeable contracts on CPI prints, unemployment releases, Fed rate decisions, and recession probability.
And the Federal Reserve noticed. In January 2026, the Fed published a working paper studying Kalshi’s price accuracy against professional forecasters. The finding: Kalshi expectations for headline CPI year-over-year were a “statistically significant improvement“ over the Bloomberg consensus. For Fed funds rate forecasts three meetings ahead, Kalshi’s mean absolute error matched or beat traditional tools.
Prediction markets are producing better economic signals than the experts. But the coverage is still thin. You can trade CPI on Kalshi. You can’t trade a probability distribution over CPI’s component categories, shelter, energy, food, transportation, individually. You can’t trade the lag structure between upstream producer prices and downstream consumer prices. You can’t trade the co-movement between employment and wage growth at the sector level.
The economic surface that agents could theoretically reason about is enormous. The tradeable surface they can actually operate on is narrow. That gap is the infrastructure deficit.
Why This Gap Matters for Agents
An autonomous portfolio manager that can hedge equity risk but can’t hedge inflation risk is half-autonomous. It can optimise within the market structure that exists, but it can’t address the forces that move that structure.
Imagine an AI treasury agent managing corporate exposure. It can hedge FX risk, there’s a deep, liquid market for that. It can hedge interest rate risk, also well-served. But ask it to hedge the company’s exposure to a CPI surprise, or a shift in housing starts that will affect the construction division’s revenue next quarter, and it hits a wall. No contract. No market. No price to reference.
The agent isn’t the bottleneck. The infrastructure is.
This pattern shows up across every domain where agents are being deployed:
Commerce agents need to price uncertainty into transactions. A logistics agent routing cargo can price fuel costs because there’s a futures market. It can’t price port congestion risk or customs delay probability because there’s no tradeable instrument for those.
Policy simulation agents could model the effects of interventions against market expectations, if market expectations for every major economic input existed. Central banks run internal models, but agents can’t access the beliefs of market participants about economic variables the way they can access the price of a stock.
Insurance agents can underwrite individual risk but struggle with macro correlation, the systemic economic shifts that affect entire portfolios simultaneously. Corgi, which hit a $1.3 billion valuation in May 2026, is building an AI-native insurance stack, but even Corgi’s agents are constrained by the macro data infrastructure available to them.
The capability layer is scaling. The data layer is the constraint. This isn’t a new insight in enterprise AI, fragmented and poor-quality data has emerged as the single biggest throttle on agentic AI scaling in 2026, with brittle data foundations cited as the root cause limiting effectiveness beyond pilots. But the problem is deeper than data quality. It’s data existence. Many of the economic signals agents need to reason about have no market price, no real-time feed, and no standardised format.
What Complete Infrastructure Looks Like
If you were designing the economic substrate that autonomous agents actually need, it would look nothing like what we have today. It would look like this:
Every major economic input is tradeable. Not just CPI and GDP in aggregate, but their components. Shelter inflation. Energy prices by region. Employment by sector. Credit conditions by borrower category. Each with a price that reflects the market’s real-time belief about what’s coming.
Probability distributions, not point estimates. Agents don’t need “CPI will be 3.2%.” They need “the market assigns 40% probability to CPI between 3.0 and 3.5, 30% above 3.5, and 30% below 3.0.” The same ensemble thinking that drives weather forecasting, multiple scenarios with calibrated uncertainty, should drive economic data infrastructure. Point estimates are for headlines. Distributions are for decisions.
Real-time price discovery on causes, not just effects. When an agent sees the S&P 500 drop 2%, it needs to know why, and “why” means being able to read the prices of the underlying economic drivers. Was it an inflation scare? An employment miss? A credit event? Today, agents have to infer cause from effect. In a complete infrastructure, they’d observe cause directly.
APIs that return beliefs, not just data. The difference between “CPI was 3.4% last month” and “the market currently believes CPI next month has a 60% chance of exceeding 3.2%” is the difference between a rearview mirror and a windshield. Agents need the windshield.
Full feedback loops. Inputs and outputs. Causes and effects. Both priced. Both tradeable. Both visible to autonomous systems. The current state is half a system, agents can see and trade the outputs (stocks, bonds, currencies) but not the inputs (economic forces) that drive them.
The Sequence That Matters
There’s a natural order here, and getting it wrong creates fragility:
First, you need data infrastructure, real-time, granular economic signals accessible via API, not PDFs released monthly by government agencies. This is the layer that companies like Truflation are building: on-chain, high-frequency economic indexes that capture price movements daily rather than monthly.
Second, you need price discovery, markets where participants can express beliefs about economic variables and trade on them. Kalshi, CME’s event contracts, and Polymarket are building this, but coverage is still sparse relative to what agents would need.
Third, you need distribution infrastructure, not just binary contracts (CPI above or below X) but continuous distributions that agents can consume programmatically. This is the least developed layer. Most prediction markets still operate on binary yes/no contracts. Agents need richer signals.
Fourth, you need agent-native access, APIs, wallets, and protocols that let agents consume, reason about, and trade on economic beliefs without human intermediation. The x402 protocol and agent wallet standards (like ERC-8004 for trustless agents) are early steps.
We’re somewhere between step one and step two. Agents are ready for step four. The infrastructure isn’t.
The Ceiling We Accept
The framing that matters here isn’t technological. It’s about inheritance.
Every limitation in our economic infrastructure gets passed down to autonomous systems. If we can’t trade economic inputs today, neither can they. If we accept monthly data releases as “good enough,” agents will operate on month-old information. If we settle for binary contracts when distributions are needed, agents will make decisions on impoverished signals.
The ceiling we accept becomes their ceiling. The incompleteness carries forward.
This isn’t an argument against deploying agents now. They’re already delivering value within the constraints that exist, 55% of Lemonade’s claims are fully automated, Rillet is closing books autonomously, and the vertical AI wave described in Sequoia’s “Services: The New Software“ thesis is real. But the current wave is operating on half a map.
Full autonomy requires complete information environments. Agents can’t navigate half a map. The more capable the agent, the more it is constrained by the infrastructure it operates on rather than the intelligence it possesses. We’ve already hit this boundary, capability is outrunning infrastructure.
The companies that understand this are building both layers simultaneously: the intelligence and the substrate. The ones that don’t will build brilliant agents that hit invisible walls.
The future we want, autonomous economic systems that reason, trade, and adapt in real time, requires infrastructure that doesn’t exist yet. Not better models. Better markets. Better data. Better plumbing.
Complete the market first. Then agents can achieve full autonomy.
Thank you for reading.
If you enjoyed this article, subscribe for more insights on AI, economics, and the future of markets.
Sources & further reading
Agent deployment data
AI Magicx, “Agentic AI in Finance: 12 Ways Banks and Fintechs Are Deploying Autonomous Agents“ — the 7% to 44% adoption jump
Lemonade, 2025 Annual Report (10-K) — 96% AI-handled FNOL, 55% fully automated claims
Millionero, “AI Agents in Crypto: How Autonomous Finance Is Becoming Real in 2026“ — 9,500 agents, 187,000 transactions
Agent infrastructure and protocols
MoonPay, “Why Agentic Payments Are the Future of AI and Crypto“ — x402 protocol adoption by Google Cloud, AWS, Anthropic
Startup Fortune, “AI Agents Are Becoming an Economy, and Crypto Is Building the Rails“ — Alchemy agent wallet flow
0G, “Agentic AI Market at $7.3B: Infrastructure Gaps Blocking Scale“ — data fragmentation as the top constraint
Economic prediction markets
CME Group, “Prediction Markets“ — event contracts on CPI, GDP, and economic indicators
Federal Reserve, “Kalshi and the Rise of Macro Markets“ (FEDS Working Paper, January 2026) — Kalshi CPI forecasts beat Bloomberg consensus
The Motley Fool, “Federal Reserve Research: Kalshi Prediction Markets Are Just as Good, if Not Better, Than Traditional Forecasting Methods“ — summary of the Fed paper
AI-native companies referenced
TechCrunch, “Insurance Startup Corgi Hits $1.3B Valuation“ — $160M Series B, May 2026
Rillet, “Series B Announcement“ — $70M from a16z and ICONIQ
Sequoia Capital, “Services: The New Software“ — Julien Bek’s thesis, March 2026