Personal AI DeFi agent: one place to design, deploy, and execute intelligent on-chain workflows through natural language.
FlowForge is a personal AI agent that makes interacting with DeFi on Arbitrum as simple as sending a message. Instead of navigating multiple dApps, dashboards, and contracts, users describe what they want in plain English, and the agent handles reasoning, workflow construction, monitoring, and execution.
FlowForge combines:
AI reasoning
On-chain execution
Oracle-based condition monitoring
Secure smart wallet infrastructure
All inside a single automated workflow.
DeFi on Arbitrum offers deep liquidity and strong yields through protocols like:
Uniswap
Aave
Compound
But usability remains a major barrier:
Users must navigate multiple interfaces.
Automation requires bots or custom scripts.
Monitoring prices and conditions is manual.
AI is not natively connected to on-chain execution.
DeFi remains powerful; but not accessible.
FlowForge turns DeFi into a conversational experience.
Users message their agent on Telegram:
“Swap my USDC to ETH if the price drops below 2000.”
The agent:
Understands intent.
Constructs a workflow.
Connects required DeFi protocols.
Monitors price conditions via oracles.
Executes the transaction automatically.
AI Layer
Intent parsing and structured action planning.
Deterministic inference powered by EigenAI.
Workflow reasoning rather than static rule execution.
Secure Execution
Runs inside EigenCompute secure enclaves.
Hardware-isolated environment.
Dedicated cryptographic identity and wallet per agent.
Wallet & Transaction Layer
Smart wallet execution via Safe.
Verifiable and non-custodial.
Oracle & Market Data
Price feeds via Chainlink.
Additional feeds from Pyth Network.
Conditional swaps
Automated lending strategies
Price-triggered reallocations
AI-driven execution flows
Telegram-native DeFi interaction
No dashboards. No manual monitoring. No scripting.
The focus of this hackathon was transforming FlowForge from a visual workflow automation platform into a fully operational AI-driven DeFi agent on Arbitrum. We implemented and deployed a production-ready agent accessible via Telegram. Key capabilities implemented: - Natural language → structured workflow translation. - Intent parsing into executable DeFi actions. - Multi-step reasoning for compound tasks (e.g., conditional swap → lending allocation). - Conditional logic support (price-based triggers using Oracle feeds). The agent is live and usable today alongside the existing web app: https://try-flowforge.vercel.app/automation-builder This is not a mock or demo - users can send real instructions and trigger real on-chain execution. Each agent instance runs with: - Hardware-isolated execution via EigenCompute. - A unique cryptographic identity per agent. - A dedicated execution environment. - Deterministic AI inference powered by EigenAI. Connected the agent identity to the user's Safe for transaction signing. This ensures: - Verifiable reasoning - Secure key isolation - Non-custodial execution - On-chain accountability The agent does not just generate text - it generates deterministic execution plans that result in signed on-chain transactions. Backend Upgrades: We extended the existing backend to support agent-driven workflows: - Updated workflow validator to handle AI-generated DAG structures. - Enhanced execution engine to support conditional monitoring and deferred execution. - Added new service routes for agent communication and execution requests. The backend now supports both visual builder workflows and AI-generated conversational workflows. We updated the frontend to support the new agent architecture: - Added onboarding flow for Safe wallet pairing. - Implemented Telegram ↔ Safe identity linking. To summarise: - Before this hackathon: FlowForge was primarily a visual Web2 + Web3 workflow automation builder. - After this hackathon: FlowForge is now a secure, deterministic, AI-driven DeFi execution agent that understands natural language.
Bootstrapped. Currently self-funded and independently developed. No external funding raised to date.
