Core concepts and definitions
This section outlines the core concepts that define AI agents in Griffin AI, rooted in foundational principles adapted from widely recognized frameworks in artificial intelligence.
“AI Agents” are digital entities with semi-autonomous capabilities, capable of executing designated tasks or services. Griffin AI develops specialized tools and frameworks tailored for blockchain applications, aimed at augmenting the autonomy of AI Agents within the blockchain ecosystem.
To guide AI agent development within Griffin AI, the following core principles (adapted from Amazon) are applied:
Sensory Perception: AI agents must be capable of sensing and perceiving their environment to gather necessary information.
Informed Decision-Making: The agents should use the data obtained from their environment to make informed decisions.
Action-Oriented: Decisions must translate into actions that are executed by the agents in the blockchain environment.
Rationality: Actions undertaken should be rational, aiming to optimize performance and achieve the most favorable outcomes.
Building on these principles, Griffin AI classifies AI agents into five distinct categories. Each category reflects a different level of capability and intelligence, catering to specific types of tasks within the blockchain framework:
Reflex Agents
Operate based on the current situation without consideration for past interactions. Actions are triggered by specific conditions in response to user events.
Exemplary task:
“Give me my current balance”, “today currency rates”
Model-based Agents
These agents incorporate an internal model of their environment that informs their decisions, allowing for responses that consider both current and past states.
Exemplary task:
”I want to buy 1 bitcoin at the price x”
Goal-based Agents
Enhance decision-making by incorporating information about desired outcomes, enabling agents to strive towards specific objectives.
Exemplary task:
“I want to invest $100 per month over the next six months in Ethereum Layer 2 projects”
Utility-based Agents
Evaluate potential actions based on a utility function, which assesses the likelihood of success and resource requirements to maximize the benefits of actions.
Exemplary task:
“I want to restake my $ETH with EigenLayer and delegate it to AVSs, with an APR of at least X% and a risk factor of at most Y%”
Learning Agents
Adapt and improve over time by learning from their environment and feedback, enhancing their decision-making and operational efficiency.
Exemplary task:
Utilize LLMs to obtain probability estimates for various events before placing bets
The concept of "Tasks" within Griffin AI is central to the functionality of AI agents. Tasks are specific assignments that outline all the necessary details required for execution, including descriptions of the task, the agents responsible, and the tools they need to employ.
Tasks vary in complexity and may involve collaborative efforts among multiple agents, often referred to as an "Agent Crew." This collaborative framework allows for the pooling of skills and resources, enhancing the efficacy and scope of the tasks undertaken. Utilizing the AI Agent Registry (see Chapter 8), AI agents are enabled to form their own crews to solve specific tasks.
An integral component of agents’ operational capabilities within Griffin AI is the array of "Tools" provided to them. These tools are designed to be multifunctional and adaptable to a wide range of needs, supporting activities across various domains. They enable AI agents to perform web searches, conduct thorough data analyses, and engage in complex problem-solving. Unique to Griffin AI are the tools provided to agents to interact with blockchain and smart contract functions. This is further explained in the chapter "Blockchain specific toolset and frameworks".
Beyond individual capabilities, these tools facilitate collaboration among agents, allowing for the delegation of tasks and the sharing of insights and results. This not only empowers AI agents to execute their designated tasks efficiently but also enhances their ability to interact and cooperate with other agents within the network.
Last updated 1 year ago.