Agent Guardian

Agent Guardian is an on-chain security layer for autonomous AI agents that need controlled access to blockchain assets.

The problem is simple: giving an autonomous agent unrestricted wallet access means a compromised key, a bad decision, or a malicious tool can turn into a direct loss of funds.

Agent Guardian puts a deterministic security boundary between the agent and the transaction.

The system has four main parts:

AgentRegistry handles agent identity, ownership, activation and recovery.

PolicyRegistry defines what an agent is allowed to do. Policies can restrict exact contract and function pairs, native-transfer targets, transaction value, daily spending and the amount that requires owner approval.

AgentExecutionGuard verifies the signed execution intent, agent status, policy, nonce, deadline, spending limits and owner approval before execution.

AgentSmartWallet keeps custody of the owner's funds. The Guard itself does not hold the user's funds. Wallet execution is bound to the specific Guard instance.

The current implementation includes EOA and ERC-1271 agent identities, exact target and selector authorization, replay protection, per-transaction and daily limits, owner approval thresholds, wallet custody, pause controls, guardian recovery and reentrancy protection.

The project is deployed on Arbitrum Sepolia.

Current verification:
178 passing Hardhat tests
95.48% statement coverage
93.63% line coverage
92.42% function coverage
77.55% branch coverage
Slither passing in CI

The current scope is intentionally focused on the deterministic on-chain enforcement layer. An off-chain AI risk engine, SDK, monitoring dashboard, production deployment and independent audit are future work.

Started with the core idea of giving autonomous agents limited blockchain permissions instead of unrestricted wallet access.

During the build, the project evolved into a full on-chain execution boundary with:

• agent identity and lifecycle controls

• policy-based authorization

• exact target and function selector checks

• transaction and daily spending limits

• owner approval for larger transactions

• wallet custody

• replay and reentrancy protection

• ERC-1271 support

• emergency pause and guardian recovery

The project was deployed to Arbitrum Sepolia and the test suite was expanded to cover adversarial execution, replay, policy ownership, wallet custody and recovery scenarios.

The current version has 178 passing tests and automated coverage and Slither checks in CI.

Not currently fundraising.

Agent Guardian is at the working prototype stage. The current focus is validating the security model, demoing the system on Arbitrum Sepolia and turning the on-chain security layer into infrastructure that other autonomous agent applications can integrate.