near-ai_skill

Overview

This skill guides NEAR AI agent development and integration for building AI-powered dApps on NEAR Protocol. It focuses on agent architecture, model integration, deployment, workflows, and security patterns to accelerate reliable agent creation. Use it to design modular agents, connect inference endpoints, and implement on-chain agent registries.

How this skill works

The skill explains prioritized rules and concrete patterns for agent architecture, AI integration, communication, model deployment, workflows, and security. It maps each area to actionable rule IDs, describes why rules matter, and provides correct vs incorrect examples for implementation. It also references NEAR AI components like the Hub, Assistant, agent registry, and decentralized inference endpoints to tie patterns to NEAR-specific infrastructure.

When to use it

• Building production or experimental AI agents on NEAR Protocol
• Integrating AI models with NEAR smart contracts or off-chain services
• Designing agent-based workflows and multi-agent coordination
• Deploying or versioning models on NEAR AI inference infrastructure
• Implementing security, privacy, and prompt-injection mitigations for agents

Best practices

• Design modular agent architecture and separate on-chain state from off-chain components
• Use structured message formats and standard agent protocols for interoperability
• Validate and sanitize inputs and outputs to prevent abuse and prompt injection
• Version and monitor models; provide fallbacks for degraded performance
• Log interactions for debugging and include human-in-the-loop checkpoints for critical flows

Example use cases

• A conversational NEAR dApp that routes user queries to specialized agents via the NEAR AI Assistant
• An on-chain agent registry that discovers agent capabilities and enforces access control
• Deploying a custom model to NEAR inference endpoints with versioned rollouts and monitoring
• Composing multi-step workflows where agents plan, execute, and persist state across transactions
• Implementing an agent gateway that validates prompts, forwards to inference, sanitizes responses, and logs outcomes