SafeAgentVault

SafeAgentVault is a control-plane plugin for agent-based DeFi systems.

It allows DAOs or protocol operators to:

• Deploy multiple autonomous agents under a shared vault
• Assign ENS identities to agents for human-readable accountability
• Enforce per-agent risk limits (caps, enable/disable, pool restrictions)
• Monitor agent runtime behavior (heartbeat, decisions, logs)

The goal is not to build a new trading strategy, but to provide a safety and permission layer around existing agents.

Starting from the practical issue of "Agents lacking controllability and security boundaries", this project is built using a modular infrastructure design. The core architecture consists of three layers: the On-chain Permission & Risk Control Contract Layer, the Agent Execution & Strategy Implementation Layer, and the Frontend Control & Visualization Layer.

In the on-chain component, we designed and implemented the SafeAgentVault smart contract, serving as the "safety exoskeleton" for Agents. The Vault handles critical risk control logics including fund custody, sub-account isolation, single-transaction limits (cap), pool whitelist verification, and Agent activation/deactivation (revoke). All swap activities must be executed through the Vault, ensuring that even if the strategy malfunctions, risks are strictly confined within calculable limits.

At the Agent Layer, we built an automated execution framework using Python, and modularly integrated existing arbitrage/sniper strategies via a unified Strategy Adapter interface. Each Agent dynamically loads different strategies based on configurations, outputs standardized execution intentions (SwapIntent) in DRY_RUN/live trading modes, and continuously writes operational status, decision rationale, logs, and heartbeat data into state.json for real-time reading by the frontend.

The frontend is constructed with React + Vite to form an interactive dashboard, acting as the "Agent Permission Console" for DAOs or strategy operators. The interface supports side-by-side display of multiple Agents, with each Agent featuring clear ENS identity, strategy type, limit, activation status, and visualization of the latest strategy decision. By polling local Agent states and monitoring on-chain events, the frontend achieves a near-real-time feedback experience.

For identity management, we integrated ENS to provide each Agent with a human-readable, auditable, and revocable on-chain identity label, significantly enhancing manageability and traceability in multi-Agent scenarios.

A noteworthy "hacker-style" design is that we deliberately decoupled strategy logic from security controls entirely. No matter how complex the strategy is, the Vault always serves as the unified permission and risk control hub—enabling the system to be reused as a general infrastructure by different AgentFi/DAO projects, rather than being tied to a specific financial strategy.