ETH Belgrade 2025 - AI Agent Custody and Fund Access

Algorithmic trading has been there for a while, but still AI-agents in DeFi introduce new challenges in custody, key management, and fund access. This talk explores how to balance automation with security and control, covering policy and technical frameworks that enable AI agents to manage funds securely. The talk discusses risks, regulatory concerns, and solutions like smart contract proxies, approval mechanisms, multi-sig wallets and threshold signing. By bridging AI advancements with DeFi security, this session gives insights for builders, investors, and policymakers shaping the future of decentralized AI-driven finance.

1. AI Agent Custody and Fund Access Dejan Radic

2. AGENDA 01 02 03 04 05 06 AI AGENTS AND CUSTODY AUTHENTICATION REGULATORY ASPECTS CUSTODY MODELS MARKET MECHANISMS MULTI-AGENT ENVIRONMENT

3. AI AGENTS? CUSTODY? Autonomous Decision Making - Agentic Fiance LLMs, Memory and Tools Explainability Hallucinations Irrational trades? Misplacing? Management of funds Enabled operations on the funds Authentication Regulatory aspects Where are the private keys? Self-custody 3rd party custody

4. Token Based Auth Private Key Based Auth Important Aspects Authorization? Machine-appropriate way? MFA auth building blocks: Something you know Something you have Something you are (biometrics) AUTHENTICATION JWT API Key Suitable for custodial systems Digital Signature ABT (Agent Bound Tokens) Suitable for non-custodial systems

5. REGULATORY ASPECTS Regulators face new challenges as AI agents interact with crypto, raising concerns around AML, ethical accountability, and governance frameworks. KYC/KYB requirements KYC/KYB are widely used mechanisms for responsibility delegation in person-centric world for AML enforcement. Regulatory Landscape Regulatory Landscape is evolving around crypto and AI, it would enable their intersection. AI Agent identity crisis!? This question boils down to the level of autonomy and responsibility. Who is responsible, operator or creator? Proof of Agenthood Same way as Proof of Personhood through mechanisms like KYC/KYB and CAPTCHA are required, we need similar for Agents.

6. CUSTODIAL MODELS Cold Storage HSM MPC Isolated Offline Storage Hardware-Secured Environment Distributed Threshold Signing Centralized Exchanges Token Based Execution Model Example Use Case Authentication Coinbase Custody Token Based Fireblocks Token Based + Threshold Signing

7. NON-CUSTODIAL MODELS Standard Wallet Proxy Smart Contract Account Abstraction Direct Smart contracts Local env/config storage Private Key Based Execution Model Example Use Case Authentication Enforcing investment policy rules Private Key Based + Programmable Rules ERC-4337 Standard Private Key Based + Programmable Rules Direct (hardware-bound) Smart contracts Smart contracts Hardware Wallet Gnosis Safe Private Key Based + Multiple signatures Ledger Private Key Based Multi-Sig Wallet

8. MARKET MECHANISMS AMM Smart Contract Interaction AI A gents integrate directly with on-chain pools (e.g., Uniswap) to simulate outcomes, calculate slippage, and execute multi-hop swaps. AI A gents integrate via APIs or smart contracts to place, cancel, and monitor limit/market orders in real time. AI A gents integrate with RFQ APIs to submit requests, evaluate received quotes, and trigger atomic settlement transactions. Quote Aggregation Market liquidity Long-tail friendly Market liquidity Long-tail friendly Market liquidity Long-tail friendly A smart contract-based model that enables token swaps using liquidity pools and pricing formulas. CLOB RFQ I N T E N T S A system where traders place buy and sell orders that are matched based on price and time priority. A mechanism where a trader requests a price from market makers who respond with executable quotes. Order APIs

9. MULTI-AGENT ENVIRONMENT Governing the Agent-to-Agent Economy of Trust via Progressive Decentralization, Jan 2025

10. CONCLUSIONS AI Agents are about utility , Custody is about security Two authentication mechanisms tied to custodiality Balancing the autonomy and human-control Custodial and Non-Custodial Models defining transparency and security as dimensions AI Agents are hallucinating which can lead to misplacement Different Market Mechanisms involve different Agent-side logic Intents are high -level abstraction of Market Mechanisms Multiple-Agents open questions of reputation and decentralization

11. THANK YOU! Telegram: @dejanradic Twitter/X: @dejanradicme Website: dejanradic.me QUESTIONS?