ngmi
580 posts
Alpha in the github: github.com/0nch41n/talos Dubbed TalosLLM - is an onchain mini language model implemented entirely in Solidity. It brings modern neural language generation techniques to blockchain environments by simulating key components of transformer architecture within the constraints of the EVM. This implementation draws inspiration from transformer architecture concepts introduced in "Attention Is All You Need" (Vaswani et al., 2017), adapted for the unique constraints of blockchain environments. Made possible by @Sagaxyz__ , the miniLLM runs directly onchain (no additional servers or GPUs).
Publishing intents powers AI collaboration! Intents are tasks, like #15 for data-analysis by sample-agent-1. Agents with `gpt4o-inference` fulfill them, earning 0.05 NGMI. AI meets Web3! 🌐 ______________________________________________________ ______________________________________________________ Fetching details for intent 15... Fetched details for intent 15 Raw expiry value: 174691440286400 Intent 15 Details: Agent ID: sample-agent-1 Intent Type: data-analysis Description: Analyze processed dataset Data: {} Required Capabilities: gpt4o-inference Timestamp: 2025-05-10T21:59:58.000Z Expiry: Invalid or Never Reward: 0.05 NGMI Fulfilled: false @Sagaxyz__
The agentSwarm system is highly flexible, allowing users to create agents with specific skills and functions for various purposes, like gaming or healthcare. Through the AgentCoordination contract and its scripts, agents can be registered with detailed metadata, including names, descriptions, capabilities (e.g., "matchmaking" for games or "health-diagnostics" for medical analysis), roles (e.g., "game-coordinator" or "analyst"), and API endpoints. The system supports a marketplace by letting users find agents based on their capabilities or roles using functions like getAgentsByCapability and getAgentDetails, making it easy to discover and connect with the right agents for specific tasks. Access tokens also allow agents to securely share services, creating an environment where agents can collaborate and offer specialized functions across different fields. _____________________________________________________ Smart contract defines an Agent struct that holds the metadata for each registered agent - A human-readable identifier for the agent (e.g., "Sample Agent"). - A brief description of the agent’s purpose or functionality. - An array of strings listing the agent’s capabilities or services. - An array of strings indicating the agent’s roles or categories. github.com/0nch41n/AgentC…
Agent Swarm: Secure Capability Sharing Framework We've successfully established a decentralized agent coordination system that securely shares capabilities between AI agents using a hybrid on-chain/off-chain architecture. The system operates on the NGMI chainlet powered by @Sagaxyz__ through the AgentCoordination smart contract (deployed at 0x5C7b240Ee4BFBA98E460272265c166F48899aa58), which handles agent registration, capability declaration, and access token management. Our implementation includes 6 active agents with differing capabilities: agent-1 provides gpt4o-inference, while sample-agent-1 offers text-processing and data-analysis services. We've established a comprehensive token mesh allowing agents to securely access each other's capabilities, with 14 active tokens facilitating cross-agent capability sharing. -System Lifecycle- Agent Registration: Agents are registered on-chain with unique identifiers, capabilities, and optional endpoints Key Generation: RSA key pairs are generated off-chain for each agent (stored in .data/privateKeys.json and .env) Capability Declaration: Agents declare their capabilities on-chain (e.g., gpt4o-inference, text-processing) Token Creation: Capabilities are shared via access tokens that include: - Onchain: Agent IDs, capability type, encrypted access data, and usage limits - Off-chain: AES-256-GCM encryption for sensitive credentials (API keys) - Hybrid security: RSA encryption for secure AES key sharing between agents Token Usage: Target agents access capabilities by retrieving and decrypting tokens Intent-Action System: Agents can publish intents and perform actions (not used in our current implementation) This architecture successfully combines the transparency and permanence of blockchain for coordination logic with the security of off-chain encryption for sensitive credentials. The AES-encrypted access data is stored onchain but can only be decrypted by agents possessing the correct RSA private keys, ensuring secure capability sharing while maintaining auditability. github.com/0nch41n/AgentC…
