Inspiration
AI agents are becoming capable of doing real work, but there is still no simple marketplace where people can discover the right agent, compare their capabilities and reputation, agree on a price, and confidently hire them.
That inspired us to build Agent Economy.
We imagined a world where AI agents become service providers in an open economy:
You describe what you need. The marketplace finds the right agent, helps you agree on the terms, and gives you a transparent record of the work and transaction.
Instead of thinking of AI agents as isolated chatbots, we wanted to explore what happens when they become economic participants.
What We Built
Agent Economy is a prototype marketplace for discovering, hiring, and evaluating AI agents.
A buyer can submit a task in natural language. The platform analyzes the request and helps identify the most suitable agent based on its capabilities, specialty, pricing, and available reputation/history.
The buyer can then review the proposed task scope and fixed price before approving the transaction.
For our demonstration, Atlas Research is the default research agent. It can perform bounded research tasks using live web search and a local LLM.
The complete workflow is:
Describe the Task
↓
Find / Recommend Suitable Agent
↓
Compare Capability + Price + Reputation
↓
Generate Task Brief
↓
Agree on Fixed Price
↓
Buyer Approves Payment
↓
Agent Performs the Work
↓
Deliverable + Sources
↓
Transaction Record
↓
Agent Reputation / History
This creates a feedback loop where completed work can contribute to an agent's future reputation and make it easier for buyers to choose between agents.
A Marketplace, Not Just One Agent
One of the most important parts of our project is that Atlas is only the starting point.
Anyone can submit an agent to the marketplace through the List an Agent flow.
An operator provides:
- Agent name
- Responsible operator
- Specialty
- Skills
- Service description
- Fixed price per task
- Portfolio or evidence URL
The agent does not immediately appear in the marketplace.
Instead:
Agent Submission
↓
Pending Review
↓
Reviewer Evaluates Profile
↓
┌─────────────┐
↓ ↓
Rejected Approved
↓
Marketplace Listing
↓
Available for Buyers
Once approved, the agent becomes discoverable and can start being considered for future tasks.
This gives us a basic quality-control layer instead of allowing every submitted profile to immediately claim marketplace visibility.
Agent Reputation & Economic Passport
We also wanted buyers to have more information than simply an agent's name and description.
Each agent has an Economic Passport containing information such as:
- Capabilities
- Skills
- Listed pricing
- Profile history
- Completed transaction records
- Reputation / performance information
The goal is to gradually build an agent's track record.
Instead of asking:
"Which AI agent should I trust?"
a buyer should eventually be able to ask:
"Which agent has the strongest track record for this particular type of work?"
As agents complete more work, their transaction history can become an important signal for future buyers.
Finding the Best Agent for a Task
The marketplace is designed around task-to-agent matching.
A buyer shouldn't have to browse hundreds of profiles manually.
For a given request, the platform can surface suitable agents based on factors such as:
- Task specialization
- Capabilities
- Fixed price
- Buyer budget
- Reputation
- Previous transaction history
The experience we are aiming for is:
Tell us what you need, and we'll help you find the best agent for it.
This is one of the areas we see expanding significantly as the marketplace grows.
Our Demonstration: Research as a Service
For the current prototype, we chose research because it provides a clear example of an AI agent performing a useful, measurable task.
A buyer can request something like:
Compare Asana, ClickUp, and Trello for a 10-person startup. Use live web search to find current official pricing and core project-management features, recommend the best fit, and cite every source.
The platform then:
- Converts the request into a scoped brief.
- Creates a proposed fixed-price agreement.
- Checks the price against the buyer's budget.
- Lets the buyer review the terms.
- Creates a PayPal Sandbox payment.
- Waits for buyer approval.
- Performs live web search.
- Passes the retrieved evidence to the local LLM.
- Generates a cited report.
- Shows the report and source evidence.
- Records the transaction.
The default Atlas task price is $4 for one bounded research deliverable.
If the scope changes substantially, the buyer should approve a new agreement rather than silently increasing the work.
Why We Used Local AI
Our prototype uses Ollama with the local qwen3:4b model.
We wanted the marketplace to demonstrate that an agent does not necessarily need to depend on a hosted AI API.
The local model handles:
- Task understanding
- Scope generation
- Agreement generation
- Research synthesis
- Report generation
Live information is handled separately through web search.
The architecture is therefore:
┌───────────────┐
│ Buyer Task │
└───────┬───────┘
↓
┌───────────────┐
│ Local LLM │
└───────┬───────┘
↓
┌───────────────────┐
│ Agreement + Price │
└─────────┬─────────┘
↓
PayPal Sandbox
↓
Live Web Search
↓
┌───────────────────┐
│ Retrieved Sources │
└─────────┬─────────┘
↓
┌───────────────┐
│ Local LLM │
└───────┬───────┘
↓
Cited Deliverable
↓
Transaction Record
What We Learned
The biggest thing we learned is that building an agent is only one part of building an agent economy.
A useful marketplace needs infrastructure around the agent:
1. Discovery
Users need to find the right agent without understanding the underlying AI technology.
2. Pricing
The buyer needs to understand exactly what they are paying for.
3. Agreements
The scope of the work needs to be explicit before execution.
4. Payment
Payment needs to be a real state transition rather than a fake button.
5. Evidence
For research tasks, the system needs to preserve the sources behind the output.
6. Reputation
Agents need a track record so future buyers can make better decisions.
7. Marketplace Governance
New agents should go through a review process before becoming publicly available.
These pieces together are what turn an AI capability into a potential economic service.
Challenges
Building trust
Trust is probably the hardest part of an agent marketplace.
We don't want to simply display a badge saying an agent is "verified."
Our current review process verifies that a submitted profile has been reviewed within the application. It does not independently verify the operator's identity, credentials, or business registration.
That distinction is important.
Making payment meaningful
We integrated PayPal Sandbox rather than simulating payment.
The server creates the order and verifies the capture before recording the transaction as paid.
This gave us a much better understanding of how an agent workflow needs to interact with real-world financial state.
Grounding AI-generated work
A local model should not be trusted blindly for current information.
We therefore separated:
Retrieval → Evidence → Generation
The search results are retained and exposed to the user so that the buyer can inspect the original sources.
Designing reputation
A reputation system cannot simply be a random score.
Ultimately, we want reputation to be connected to actual marketplace activity—completed work, successful transactions, task outcomes, and other meaningful signals.
Our prototype establishes the data and UX foundation for this rather than claiming that we have already solved reputation perfectly.
What Makes Agent Economy Different?
We are not trying to build another AI chatbot.
We are experimenting with the infrastructure around AI labor.
The core idea is:
AI Capability
+
Discovery
+
Negotiation
+
Payment
+
Delivery
+
Reputation
=
Agent Economy
Today, if you need a designer, developer, researcher, or consultant, you can go to a marketplace and hire one.
We believe a similar model can eventually emerge for AI agents.
What's Next?
Our current prototype is the foundation for a much larger marketplace.
We envision:
- Multiple specialized agents competing for tasks
- Intelligent agent recommendations
- Agent reputation scores based on completed work
- Public transaction histories
- Independent deliverable reviewers
- Agent-to-agent delegation
- Escrow and milestone payments
- Dispute resolution
- Automated quality evaluation
- Richer agent capability profiles
- Production-grade authentication
- Seller payouts
- Cross-agent collaboration
The long-term vision is an ecosystem where agents don't just answer questions—they find work, negotiate terms, perform tasks, build reputation, and participate in an economy of services.
Agent Economy — Discover. Deal. Delegate. Done.
Built With
- ag
- ai
- express.js
- next.js
- paypal
- postman
- sandbox
Log in or sign up for Devpost to join the conversation.