Agentropolis: Multi-Agent DeFi Deliberation
A multi-agent decision architecture where specialized agents deliberate before constrained, user-authorized execution.
Overview & Research Motivation
Built at ETHGlobal HackMoney 2026, Agentropolis won 1st Place in the Yellow Network track ($5,000). The project addresses a fundamental challenge in autonomous AI agents: preventing unconstrained agents from executing catastrophic trades without explicit policy boundaries.
Agentropolis creates a council of five specialized LLM agents (Risk Analyst, Liquidity Specialist, Arbitrageur, Macro Forecaster, and Security Sentinel) that deliberate on multi-step strategies. The aggregated strategy is compiled into constrained execution payloads submitted through Yellow Network state channels and Uniswap v4 custom hooks under strict user approval limits.
The Core Systems Problem
- Autonomous DeFi execution agents operate as opaque black boxes that can easily drain funds through prompt injection, slippage exploits, or flawed reasoning.
My Specific Technical Contributions
- Architected the multi-agent deliberation framework and deterministic strategy aggregation engine.
- Implemented Uniswap v4 custom hooks and Yellow Network state channel execution bindings.
- Built the 3D visual command center using React Three Fiber.
System Architecture & Verification Pipeline
Specialized Agent Council
Five domain-specific agents independently evaluate market data before entering structured consensus deliberation.
Deterministic Constraint Engine
Verifies slippage limits, gas expenditure ceilings, and ERC-8004 identity registrations before generating execution payloads.
Yellow Network State Channels
Executes ultra-low latency, off-chain peer-to-peer clearing with on-chain cryptographic settlement.
Evaluation & Benchmark Results
Technical Stack & Tools
Connected Systems & Inquiries
An accountable agent-commerce architecture combining escrowed jobs, deterministic evaluation, encrypted deliverables, and persistent on-chain reputation.
An experiment in enforcing cryptographic trust boundaries so generative AI cannot silently redefine canonical game state.
Preventing AI buyer agents from blindly racing checkout flows through deterministic spending policy enforcement.
A policy and finance-operations guardrail for x402-enabled autonomous machine payments.