How Do Autonomous Economic Agents Work?

Noone21 is creating AI agents equipped with Autonomous Economic Agent (AEA) capabilities. AEAs are divided into two categories: Passive AEAs and Active AEAs. Users send compensation and costs required for task execution to the AEA when requesting a task. The AEA processes the request, completes it, and either settles accounts or pays other AEAs as needed. In this way, each AEA becomes an independent participant in economic activities.

 

• Passive AEA: Executes tasks using pre-defined processing workflows that incorporate tools like LLMs (Large Language Models), memory, ontology, and other resources to derive results.
• Active AEA: Actively adapts to external feedback using LLMs until reaching the final conclusion.

The following is the Noone21 AEA Reference Model designed for KLEVA DeFAI.

Noone21 KLEVA AI Agent Reference Model

 

Here’s a simplified explanation:

1. User Request: The user sends a specific request to the AEA.
2.System Prompt Creation: The AEA analyzes the request and generates an enhanced system prompt to guide internal processing. This prompt is then passed to the Reasoning Loop.
3.Reasoning Loop: This loop iteratively processes the task by incorporating feedback from the “external environment” and refining its reasoning until a final answer is derived.

Reasoning Techniques:

• ReAct: Combines reasoning and action, where the agent alternates between breaking down the problem (reasoning) and interacting with external tools or data sources (action).
• CoT: Guides the agent to explain intermediate reasoning steps for complex tasks.
• ToT: Structures reasoning as a tree, exploring multiple branches of thought to find optimal solutions.

 

1. User Request:
• “Please sell all my KLEVA tokens when the price is about 50% higher than the purchase price.”

2. ReAct Process:
• Reasoning (Thought): “To sell KLEVA tokens, I need the selling platform, the seller’s wallet and signature, and the transaction fee.”
• Action:
Ask the user: “Could you tell me the purchase price of your KLEVA tokens and transfer a transaction fee of 0.001 ETH to my wallet?”
• Observation (Action_Response):
The user replies, “1,000 Won,” and transfers 0.001 ETH as the transaction fee to the agent’s wallet.
• Further Reasoning (Thought):
“To sell KLEVA tokens, I need to access multiple cryptocurrency trading platforms and get the user’s signature when the price reaches ‘1,500 Won’.”
• Action:
Access multiple trading platforms, monitor when the price increases by 50% above the average purchase price, and request the user: “Could you provide your signature for the sale?”
• Observation (Action_Response):
The user provides their signature.
• Final Output:
“The sale of your KLEVA tokens has been completed.”

(4) During the Reasoning Stage:  If necessary, the AEA sends prompts to the LLM (Large Language Model) to perform specific tasks.
(5) External System Calls:
The AEA can call Extensions, Functions, or use Retrieval methods to access Datastores when external system interactions are required. For example, in the case of accessing multiple trading platforms to check market prices and execute sales, predefined Extensions and Functions are utilized.
(6) Context Management and Memory Usage:
Throughout the reasoning process, the AEA maintains context and stores information related to queries (prompts) using short-term and long-term memory. To continuously refine prompts, Ontology is employed to provide synonyms and related terms, enabling richer and more accurate responses.

A key feature of AEAs is their ability to adapt actively to external environments. For instance, if monitoring Trump’s X (formerly Twitter) account—which significantly influences cryptocurrency prices the AEA could detect news about Trump’s strategy to stockpile crypto assets. Based on this information, it might notify the user and suggest adjusting the selling price threshold from 50% to 60%.

 

Economic Activities of AEAs:
Since every AEA agent possesses its own wallet account, it can engage in various economic activities. These include receiving compensation and costs for task execution or providing rewards to users based on completed tasks.