Google DeepMind: The economic layer where AI independently creates value is taking shape.
Is the AI Agent giving rise to a brand - new economic layer?
Google DeepMind and the University of Toronto jointly stated: Yes.
Here, agents can conduct transactions and collaborations on a scale and at a speed beyond direct human supervision.
To conduct an in - depth analysis of this emerging system, they also jointly proposed the “Sandbox Economy” and characterized the new economy from the following two key dimensions:
Origin (spontaneous emergence vs. human - designed);
Degree of separation from the existing human economy (permeable vs. impermeable).
Besides, Nenad Tomasev, a senior research scientist at DeepMind, also said:
The rapid popularization of AI Agents indicates that a new economic layer where AI independently creates value is quietly taking shape.
So, how does this brand - new AI agent economy operate, and what opportunities and challenges will it bring?
Let's find out.
A Comprehensive Analysis of the New Economic Layer of AI Agents
With the continuous explosive growth of agent systems and the continuous development of new operating standards such as the MCP protocol, the emergence of a new economic form is inevitable.
This emerging system can be called either the “virtual agent economy” or the “sandbox economy.”
It's worth mentioning that the name “sandbox economy” especially reflects its core goal - to ensure that AI agents can operate safely and reliably within the economic layer.
Judging from the current development trend, we are moving towards a spontaneously emerging and highly permeable AI agent economic system.
Next, let's intuitively demonstrate the specific applications of AI agents in economic activities through three typical scenarios.
Scientific Research
For example, under the guidance of a mathematics professor, GPT - 5 extended the qualitative fourth - moment theorem to a quantitative form with an explicit convergence rate for the first time.
The AI Agent named Gauss completed the mathematical challenge proposed by Terence Tao and Alex Kontorovich - formalizing the strong prime number theorem in Lean - in just three weeks.
Robotics
In terms of housework, robots can help us do the laundry, wash the dishes, wipe the tables...
They can also work in factories and sort express deliveries.
Personal Assistants
For example, Meituan's newly launched Agent Xiaomei allows people to order takeaways just by speaking.
There are also many office assistants that can directly help us organize materials and generate reports, improving work efficiency.
However, when multiple agents interact on behalf of different users simultaneously, conflicts are inevitable - for example, what should we do if everyone wants to grab the same resource?
In response to the above problems, researchers proposed to arrange resources by means of market mechanisms and fair distribution rules.
Imagine a virtual market: each user's AI agent has the same amount of “virtual currency” to bid for shared resources, such as computing power, data access rights, or opportunities to execute tasks with priority.
In this way, each agent is fair in terms of ability, and users will not suffer losses due to differences in agent strength.
The prices of resources are naturally formed through agent bidding, and resources ultimately flow to the places where they are most needed. At the same time, the resource combination obtained by each agent is as consistent as possible with the user's preferences, ensuring that no one will envy others' resources.
This design not only reflects users' needs but also prevents unfairness caused by uneven abilities, laying the foundation for a controllable and fair AI sandbox economy in the future.
Effective Supervision is Required in Multiple Fields
Besides, to achieve a practical and safe virtual agent economy, people also need to take targeted measures in multiple fields such as law, technology, and policy.
Clarify Legal Responsibilities: Who Should Pay for AI's Actions?
Break through the traditional “single - subject liability” model, regard the multi - AI collaborative system as a “collective liability entity,” and clarify the responsibilities of AI creators (such as technology companies), deployers (such as enterprises), and users (such as individuals) in different scenarios.
For example, if an AI causes trading losses due to algorithmic defects, the creator needs to bear technical responsibility; if a user maliciously instructs an AI to commit fraud, the user needs to bear the primary responsibility.
Unify Technical Standards: Avoid “AI Language Barriers”
Promote the popularization of interoperability standards such as the A2A protocol and the MCP protocol, so that AIs from different companies and different fields can “communicate smoothly” and avoid the formation of a fragmented ecosystem like a “walled garden.”
Build a “Three - Level Supervision” System: AI Supervises AI, and Humans Provide a Safety Net
The first level is for “supervisory AIs” to monitor the market in real - time and intercept fraudulent and abnormal transactions;
The second level is to quickly contain hazards through automated protocols (such as freezing the accounts of violating AIs);
The third level is to transfer complex cases (such as large - scale economic losses caused by AIs) to human experts for handling. At the same time, use an immutable ledger to record all transactions to ensure that problems can be traced and reviewed.
Conduct “Regulatory Sandbox” Pilots: Take Small Steps and Validate Gradually
Jointly with enterprises, research institutions, and regulatory authorities, test small - scale AI economies in specific scenarios (such as energy optimization on university campuses and autonomous driving delivery in cities), observe the collaborative/competitive behaviors of AIs, and evaluate fairness mechanisms.
Protect Human Values and Promote “Human - AI Collaboration”
On the one hand, reform the education system to strengthen human advantages in areas such as critical thinking, complex problem - solving, and AI output evaluation (for example, train “AI coaches” instead of letting AI completely replace human decision - making).
On the other hand, improve the social safety net. In response to the possible replacement of the labor force by AI, strengthen unemployment protection and portable benefits to ensure that the wealth created by AI can benefit everyone, not just a minority.
One More Thing
Today, the world's first AI Agent trading market, MuleRun, is officially launched and open to all users. It is also the world's first AI worker marketplace, that is, the AI digital labor market.
Through a unified entrance, MuleRun aggregates a wide variety of agents, global creators, and a large number of users, realizing one - stop agent trading services.
Moreover, MuleRun has also launched the Global AI Agent Creator Support Program, providing cash support, marketing support, and technical support to accelerate the cold - start process of creators and helping creators achieve long - term and stable income growth through continuous incentives.
Since there is a market for agents, it seems that we are really not far from “winning effortlessly” at work.
Reference Links:
[1]https://arxiv.org/abs/2509.10147
[2]https://x.com/weballergy/status/1967478089018908748
[3]https://x.com/_akhaliq/status/1967579770075627587?s=46
This article is from the WeChat official account “QbitAI”. Author: Shiling. Republished by 36Kr with permission.