ActionVM
Last updated
Last updated
ActionVM is the foundational component of the Magnet Protocol, designed to provide a robust runtime environment for Action Agents. It features the following key capabilities:
Multi-LLM Support: ActionVM is compatible with various large language models (LLMs), including ChatGPT, Claude, LLaMA 3, and Mixtral, offering flexibility and scalability for different use cases.
Unified Runtime for Creator Hub: It acts as the runtime for standardized Action templates, enabling the Creator Hub to streamline the creation and customization of Action Agents. Users can configure templates easily to suit specific needs without technical expertise.
Open-Source and Decentralized: ActionVM is fully open-source and designed to support a decentralized network. In the future, node participation will be gradually opened, ensuring a distributed infrastructure for running Magnet Action Agents securely and efficiently.
These features make ActionVM the backbone of the Magnet ecosystem, bridging LLM capabilities with decentralized agent execution.
As illustrated in the diagram, ActionVM comprises the following key components:
The Action Assembler (AA) is responsible for loading all Actions associated with an Action Agent, serializing them, and preparing the necessary foundational service dependencies, such as Twitter API, Farcaster API, Coinbase CDP, or AWS Bucket. It interacts with the LLM Linker to wait for its results. Upon successful serialization, the Action Agent is passed to the Runtime for execution.
AA also serves as the template validation tool for the Creator Hub, ensuring the serialized correctness of generated Action Agents. Essentially, AA functions as the “compiler” of ActionVM, converting static Action Agent definitions into executable versions.
The LLM Linker (LL) initializes API connections to various large language models (LLMs) based on configuration data provided by AA. Using the Service Key Manager (SKM) within a Trusted Execution Environment (TEE), it retrieves the necessary API access keys for LLMs.
LL supports major commercial models like ChatGPT and Claude, open-source models like Hugging Face Transformers, and custom models hosted on major cloud platforms such as AWS SageMaker and Azure ML.
The Service Key Manager (SKM) handles the secure management of service keys required by Action Agents, such as Twitter API keys, Farcaster credentials, private keys, and Coinbase CDP developer keys. Using TEE technology, SKM ensures all keys are securely stored and accessed.
Knowledge Storage (KS) is divided into two segments:
Private Knowledge: Stores data essential for private Action Agents, including knowledge streams and user-defined data during creation.
Public Knowledge: Maintains data collected by the Agent during runtime, such as social data, on-chain data, and centralized exchange (CEX) data.
Private knowledge is encrypted using TEE technology, similar to SKM, ensuring secure storage of all sensitive information.
The AI Gas System (AGS) calculates token consumption between different Action Agents and the LLMs they interact with. It forwards the AI Gas consumption results to the Agent Marketplace for cost tracking and billing purposes.
The Runtime is the core operational component of ActionVM. It loads and executes serialized Action Agents by running their associated code programs, ensuring seamless execution.
These components collectively enable ActionVM to provide a secure, scalable, and modular environment for running sophisticated on-chain AI agents.
