Agent Guardrails Docs

Agent Guardrails Protocol

Agent Guardrails is an on-chain policy layer for AI agents operating on Solana. It sits between autonomous agents and the blockchain, enforcing allow-lists, spending budgets, and a real-time AI kill switch so that operators never lose control of their deployed agents.

The protocol intercepts every agent transaction through a single entry point (guarded_execute), validates it against three layers of defense, and can freeze a compromised agent on-chain in under 3 seconds. Funds live in a program-derived address controlled by the policy — never in the agent keypair — so even a fully compromised agent cannot drain its treasury.

Three Layers of Defense

  • Layer 1: Program Allow-Listing— Only whitelisted programs (Jupiter, Marinade, Drift, etc.) can be called via CPI. Any instruction targeting an unknown program is rejected on-chain before execution.
  • Layer 2: Spending Budgets— Per-transaction caps and rolling daily budgets enforced on-chain. The SpendTracker PDA tallies every outflow and rejects transactions that would exceed limits.
  • Layer 3: AI Kill Switch (Guardian Agent)— An off-chain monitoring pipeline analyzes every transaction for behavioral anomalies. When the Guardian Agent detects a threat, it signs a pause_agent instruction that freezes the agent on-chain immediately.

Repository Structure

DirectoryDescription
program/Anchor 0.30.1 Solana program — on-chain policy enforcement, guarded execution, and pause/resume instructions
server/Express API + worker pipeline — webhook ingestion, Guardian Agent judge, incident reporting, SSE push
dashboard/Next.js 14 frontend — real-time monitoring interface with live activity feeds, spend gauges, and incident timelines
sdk/TypeScript client and IDL — source of truth, synced to server and dashboard via build script

Who It's For

  • Autonomous agent operators— individuals or teams deploying AI agents for trading, yield farming, or portfolio rebalancing who need guardrails around what those agents can do.
  • Institutional operators (DAOs, funds)— organizations that require policy enforcement, audit trails, and multisig escalation before any high-value agent action.
  • AI-native protocols (Jupiter, Marinade, Drift)— protocols that want to offer agent-friendly integrations with built-in safety guarantees for their users.

Built With

  • Solana — high-throughput L1 for sub-second finality
  • Anchor 0.30.1 — Solana program framework for account validation and CPI
  • Helius — real-time webhooks for transaction monitoring and enhanced RPC
  • Squads v4 — multisig escalation for high-severity incidents
  • Guardian Agent — LLM-powered anomaly detection and behavioral analysis
  • Neon Postgres — serverless Postgres for transaction history and incident storage
This project was built for the Solana Frontier Hackathon.

Quick Links