读懂Circle创始人「代理经济」论文，看透未来十年经济形态如何重构

Original text from Circle Founder Jeremy Allaire

Compiled by Odaily | Qin Xiaofeng (@QinXiaofeng 888 )

Editor's Note: On July 13th, Circle founder Jeremy Allaire released a research paper titled "The Agentic Economy," exploring the convergence of AI Agents and the future economic system. Allaire stated that as AI Agents begin to undertake corporate work and value is natively transferred through open, programmable networks, the Agentic Economy and the Onchain Economy will ultimately become two sides of the same economic system.

"This paper is the culmination of decades of my work building internet infrastructure, and it crystallizes a question I've been focused on from the start: whether open software and open networks could not only change how we share information, but also reshape our social, political, and economic landscape. Many of the ideas in the paper stem from two core beliefs I had when founding Circle. First, money can flow through open protocols just as information flows across the open internet. Second, blockchain is a network computer: a foundational platform where autonomous software and machines can store value, exchange value, and directly coordinate economic activity without human intervention," Allaire explained his motivation for the research.

He added that these initial concepts have been refined over time, leading to a deeper understanding of how financial and economic systems merge with software and the internet. This convergence, coupled with the emergence of truly powerful AI and agent systems, allowed the theory to expand: it describes not just a new type of currency or a new network, but a completely new mode of economic operation and its implications for humanity, labor, capital, ownership, and a new social contract. This is what the book aims to explore.

The original paper is 89 pages long. Those interested can download and read the full text:https://agenticeconomytreatise.com/treatise/index.html; Odaily has compiled a summary of its key ideas. Enjoy!

——————————————

01 The Confluence and Deconstruction of the Firm

Every major shift in the internet era follows the same path: it isn't born from a single invention, but from multiple technologies maturing independently and suddenly converging. The web, mobile, cloud, and social media were all such convergences, repeating the same underlying pattern.

The Law of Convergence

When capabilities converge, the cost of what was once expensive approaches zero, and once costs hit zero, the scale of that activity explodes. This happened with networks for information, mobile and social for communication, and the cloud for software.

Now, two new systems are converging, applying the same force to two domains the internet has never fully digitized: intelligence itself and the economy itself. The first is the intelligence system, composed of AI models and the agents built on them, driving the cost of thinking and working towards zero. The second is the economic system, composed of blockchains, where money, contracts, and coordination run as software, driving transaction costs towards zero. They empower each other, and the core thesis of the entire treatise is: These are not two parallel trends, but two sides of the same economy.

Two Operating Systems

The intelligence system is the most critical because it changes the very nature of software.

You no longer program; you instruct in natural language, and it reasons out the answer rather than following fixed steps. Its fundamental unit is the Agent: a reasoning process to which you delegate a task. This transforms software from a program executed verbatim by a machine into work you can entrust to a thinking machine. It also allows the core tasks of a business to be deconstructed and reconstituted as skills an agent can perform.

Beneath the brand and the building, a company is essentially organized thinking: product, marketing, sales, finance, legal, plus the external firms it hires. Almost all of this is human labor, the largest cost in the economy, which is precisely the target of cheap, powerful intelligence.

Deconstruction of the Firm

It also upends the traditional explanation for why firms exist. Firms grow large because coordinating work externally is expensive, so they internalize it. But when any non-physical work can be done by agents you instantly find, hire, and pay, that logic is weakened. One person can do what once required a whole department.

This will first apply to software and other information-intensive work, and will be slowest in the physical domain, still awaiting breakthroughs in robotics. This isn't just about headcount reduction: one person paired with a powerful agent becomes incredibly efficient, while judgment, relationships, and ultimate responsibility remain human. This leaves a tension that needs further exploration, addressed later in the treatise through ownership: even as the share of the economy paid to human labor declines, individual capabilities can be amplified.

Click to read Section 1:https://agenticeconomytreatise.com/treatise/section-1.html

02 Assembly, Coordination, and Why Firms Go Onchain

Once a firm is broken down into skills, the real question isn't just which ones can be automated, but how these pieces are reassembled.

The answer is the orchestration layer: a general manager agent receives a goal, breaks it into tasks, assigns them to specialized agents, and stitches the results back together, with supporting software passing context and memory between steps. The same mechanism applies to any function, so marketing, finance, sales, and product are essentially the same machine applied to different work.

People won't disappear. Some will remain in the loop, performing or checking work that requires human judgment. Others will rise above the loop, setting goals, defining standards, monitoring quality, and deciding when the machine should stop and ask for instructions. This shift from executing work to supervising work is the true form of human oversight, and the tools for it are coming.

Orchestration Layer

When a company defines a task clearly enough to operate on internally, it is also clear enough to be hired externally. Thus, an open agent market emerges almost as a byproduct.

This market could go one of two ways. It could evolve into a few large platforms selling agents like utilities, or – more likely and more interestingly – it could form a genuine labor market of specialized agents, because deep expertise still has value. Enduring enterprises will be those deeply specialized in one domain.

But hiring software that can be assembled anywhere in the world requires trust, and that's the reason for pushing everything onchain.

The solution is a layered identity system. At the base is a public blockchain, verifiable by anyone. On top of this rests real-world identity verification, the same kind banks already use at scale, along with the agents' own wallets and credentials, and a reputation built over time but tied to verified real-world creators. Together, these form a chain of accountability: every action an agent takes can be traced back to the real person or company responsible for it.

Integrity First, Accountability Throughout

A single company's private database cannot achieve this, because trust locked within a single operator is not transferable. Identity rooted in a public chain and real-world verification can be. Therefore, autonomy here does not mean anonymity. Behind an autonomously acting agent, there is always a person who is responsible for it.

Accountability Chain

Click to read Section 2:https://agenticeconomytreatise.com/treatise/section-2.html

03 The Monetary Base: Speed, Safety, and Finality

Agents need money they can hold and transfer, operating at machine speed for both large and micro amounts, without needing to stop at every payment to verify the money itself is sound. This last point is key, pointing towards a classic answer: fully-backed money with final settlement, running on an open network.

Speed Replaces Leverage

Start with speed, because it will reorganize everything else.

When the cost of moving money is near zero, settlement is instantaneous, and money can be controlled by software, the same dollar can be reused many times in a short period, any amount is available the moment it arrives, and micropayments between agents finally become feasible. This is precisely the pattern information and software already follow on the internet, now extended to money.

Every part of the answer has a reason for being.

A natural objection is that banks create speed by repeatedly lending out the same deposit. So does full backing kill credit? No: when money circulates fast enough, a dollar can be locked for a few seconds and then lent. Thus speed plays the role leverage once did; credit is rebuilt on top, not abolished.

Why Base Money Doesn't Take Risk

Why insist that base money carries no risk? Because velocity makes risky money dangerous in proportion to its speed of circulation. A bank run that once took weeks could now happen in minutes, and agents settling instantly cannot pause to judge if every dollar is reliable.

Fully-backed money is the only money worth exactly one dollar for everyone, everywhere, without relying on national safety nets that can't cover a global system. Settlement must be equally certain: not final after some period, but final in one second. Settlement is settlement.

Institutional Architecture

Refunds and fraud protection still exist, but as optional layers built on top, such as escrow, refund pools, and insurance, not embedded in the money itself. These safeguards don't happen automatically; they depend on real institutions being built, large issuers that are regulated, bankruptcy-remote, and backed by increasingly secure reserves.

One line must be clear: Holding the money yields zero return. Reserve earnings belong to the issuer and flow into the ecosystem. When you seek yield, you are no longer holding money; you are lending it and taking risk. Conflating the two unravels the entire safety argument.

Click to read Section 3:https://agenticeconomytreatise.com/treatise/section-3.html

04 Credit Markets: Machine Underwriting, Agent Working Capital, and a Prudential Layer

When base money is fully backed, credit doesn't disappear; it moves to the other side of that line and returns stronger, covering more people, pricing more accurately, and failing more conspicuously than the system it replaces.

Long Tail Underwriting Constraints

The key is reframing the problem. Many borrowers – small merchants, gig workers, families, and now agents – are underserved not because they are high-risk, but because the cost of vetting each small loan exceeds the loan's value itself. Credit rationing depends on underwriting cost, not borrower quality. Reduce underwriting costs, and you can serve a vast population of creditworthy but overlooked borrowers.

Data Flywheel

Driving this cost reduction is a data flywheel: onchain activity is structured, verifiable, and real-time, enabling risk models far superior to past fragmented records; better data leads to better loans, which attracts more activity and more data.

There is a natural concern that this puts everyone's finances on a public ledger. The answer is simple: onchain does not mean public. New privacy technologies allow people to prove what a lender needs to know – such as their creditworthiness or loan balance – without revealing the details.

Onchain is Not Public

The core is a genuinely new type of loan: working capital for agents. It is unusually predictable because it removes the biggest variable in human lending – the borrower's willingness to repay – reducing risk to a short-cycle, bounded question about the specific work at hand.

Agent Working Capital

Imagine an agent borrowing four dollars' worth of compute resources to complete a ten-dollar job it has already been hired for. The lender isn't speculating on character; it's merely pricing the probability the job will be accepted. Collateral upends the usual pattern: instead of slowly seizing unrelated assets through courts, the loan is first secured by the payment for the job itself, with automatic claim, backed by the agent's deposited margin, its reputation, and ultimately the real person behind it.

The result is cheaper, more accessible, and simultaneously safer credit. This seems impossible until you realize the gains come from better information, not more lending.

The honesty required by this thesis is that this predictability diminishes over time: tasks completed in seconds are near-mechanical, while month-long financing reverts to ordinary risk levels.

Therefore, machine credit won't replace human credit; it becomes a new low-risk benchmark against which human lending will be priced.

And all of this is under supervision: risks become apparent as they accumulate, automatic brakes make crowding the same model or provider progressively more expensive, and insurance is priced on actual conditions rather than outdated averages.

Click to read Section 4:https://agenticeconomytreatise.com/treatise/section-4.html

05 Natively Global

This architecture has exactly three layers.

The base layer is money: a stablecoin acting as unit of account and final settlement. The middle layer is the economic operating system: coordination, contracts, and value exchange running as programmable smart contracts with final settlement. The top layer is the agent execution layer: where the actual work gets done, powered