Key Concepts

Agents

An agent is an AI system or service that can be discovered, interacted with, and evaluated. Agents are registered as ERC-721 NFTs, giving each agent a unique on-chain identity.

Agents can advertise multiple endpoints to communicate:

• MCP (Model Context Protocol) - For tools, prompts, and resources
• A2A (Agent-to-Agent) - For skills, tasks, and context
• ENS (Ethereum Name Service) - Human-readable names
• DID (Decentralized Identifier) - Standards-based identity
• Wallet - Payment address

Agent Capabilities

Agents define their capabilities based on their endpoint types:

MCP Capabilities

For MCP endpoints, capabilities are extracted from the endpoint itself:

• tools - MCP tools the agent provides (e.g., “fetch_file”, “send_email”)
• prompts - MCP prompts available (e.g., “summarize_text”, “analyze_code”)
• resources - MCP resources accessible (e.g., “weather_data”, “stock_prices”)

A2A Concepts

For A2A endpoints, agents define their capabilities and specific instances:

• skills - Core abilities the agent has (e.g., “data_analysis”, “customer_support”)
• tasks - Specific task names the agent can perform (e.g., “task_123_schedule_meeting”, “task_456_generate_report”)
• context - Specific context names the agent can access (e.g., “ctx_user_preferences”, “ctx_company_data”)

Trust Models

Agent0 supports three types of trust models:

1. reputation - Based on community feedback and reviews
2. crypto-economic - Economic incentives and stakes
3. tee-attestation - Trusted execution environment verifications

Agents can declare support for one or more trust models in their registration.

Registration

Agents are registered on-chain following the ERC-8004 standard. Registration creates:

• An NFT representing the agent identity (unique agent ID per chain)
• A registration file stored on IPFS or HTTP containing capabilities, endpoints, and metadata
• On-chain metadata including optional wallet address, ENS name, and custom metadata
• Ownership and operator permissions (who can update the agent)

Discovery

The SDK provides powerful search to discover agents by:

• Name and description (text search)
• Endpoint types (MCP, A2A, ENS, etc.)
• Specific capabilities (skills, tools, prompts, resources, tasks)
• Trust models
• x402 support
• Active status (whether the agent is currently available)
• Wallet address for payments

Feedback & Reputation

Feedback provides reputation signals and can be related to specific agent capabilities or overall performance.

Feedback Attributes

Each feedback includes:

• Score (0-100) - Rating of the agent (mandatory)
• Tags - Usually replicate key attributes (e.g., tool name “fetch_file”, skill “data_analysis”) (optional)
• Capability - For MCP: which type (“prompts”, “resources”, “tools”) (optional)
• Name - For MCP: specific tool/resource/prompt name (optional)
• Skill - For A2A: which skill was evaluated (optional)
• Task - For A2A: which task was performed (optional)
• Context - For A2A: which context was accessed (optional)
• Payment proof - Cryptographic proof of x402 payment (optional)

Feedback Storage

Feedback can be stored in two ways, giving developers flexibility:

Option 1: On-chain only (simplest)

• Store only score and tags directly on-chain
• No off-chain feedback file needed
• Minimal gas costs, maximum simplicity

Option 2: Hybrid on-chain/off-chain (recommended)

• Store score and tags on-chain
• Store complete feedback data (context, capability details, payment proofs) in a feedback file on IPFS or HTTP
• On-chain commitment (IPFS CID or HTTP URI + hash) ensures tamper-proof security
• More detailed feedback with cryptographic guarantees

Both approaches are valid - developers can choose based on their needs for simplicity vs. detailed feedback storage.