PRAETOR_AGI: Autonomous Treasury Sentinel

dashboard
mnee

Inspiration

We are witnessing the dawn of Agentic Finance. AI models are getting smarter, but they remain "economically paralyzed"—they can write code or generate art, but they cannot pay for a server, buy an API key, or invest in liquidity pools.

The problem isn't just technical; it's psychological. Giving an autonomous agent access to a wallet is terrifying. The "Black Box" problem means we don't know why an agent is spending money until the funds are gone.

We were inspired to build PRAETOR_AGI to solve this "Trust Gap." We wanted to create a system where an AI could be an autonomous CFO—capable of earning and spending the MNEE stablecoin—while operating inside a transparent "Glass Cockpit" where a human can watch, understand, and override every decision in real-time.

What it does

PRAETOR_AGI is an autonomous treasury sentinel that interfaces directly with the Ethereum blockchain. It transforms natural language directives into on-chain financial actions using the MNEE stablecoin.

Autonomous Reasoning: You give it a high-level goal (e.g., "Write and sell a profitable article"). The agent (powered by Google Gemini 2.5 Flash) analyzes the request to calculate the estimated cost, potential revenue, and risk level.
The Entrepreneur Loop: Unlike simple "spender" bots, PRAETOR can identify ROI. It will spend MNEE to acquire resources (APIs, Data) to produce assets that generate more MNEE, effectively running a profitable autonomous business.
Glass Cockpit Interface: Every thought, API call, and transaction hash is streamed to a live dashboard. The user is never in the dark.
The "Kill Switch": A physical-style ABORT button allows the human operator to sever the agent's autonomy instantly if it detects high-risk behavior or hallucinates.
Guardrails: The system automatically rejects requests that violate ethical guidelines or exceed spending limits, ensuring the treasury is never drained by malicious prompts.

How we built it

We built PRAETOR_AGI using a modular architecture designed for speed and security:

Brain (AI): We used Google Gemini 1.5 Flash for its low latency and superior reasoning capabilities. It acts as the "CFO," returning structured JSON data regarding Risk (), Cost (), and Urgency ().
Body (Blockchain): We utilized Web3.py to interface with the Ethereum Mainnet. The system interacts directly with the MNEE Token Contract, managing allowances and executing transfer or withdraw functions.
Nervous System (Backend): A FastAPI server handles the asynchronous state management, bridging the gap between the slow, deterministic blockchain and the fast, probabilistic AI.
Eyes (Frontend): A React dashboard serves as the "Glass Cockpit," visualizing the ledger and logs in real-time using high-contrast, data-dense UI principles.

Challenges we ran into

The "Hallucinating CFO" Problem: Early on, the AI would happily approve spending 1,000,000 MNEE for a "coffee." We had to engineer strict system prompts and JSON validation to force the model to respect the value of money.
Async State Management: Syncing the AI's "thinking time" with the blockchain's "confirmation time" and the frontend's "rendering time" caused race conditions where the "Abort" button wouldn't work fast enough. We solved this by implementing multiple await checkpoints in the Python backend loop.
Mainnet Simulation: Testing on Mainnet is expensive. We built a robust "Simulation Mode" that mimics Mainnet latency and contract interactions, allowing us to demonstrate the full flow without burning real gas during development.

Accomplishments that we're proud of

Real-Time Guardrails: We successfully implemented a logic gate where the AI refuses to spend money on illegal or high-risk requests (like "buying malware"), purely based on context analysis.
The "Profit Logic": Teaching the AI to distinguish between an Expense (losing money) and an Investment (spending to earn).
Seamless Integration: Successfully bridging the gap between a Generative AI model and a deterministic EVM Smart Contract.

What we learned

Transparency Creates Trust: Users are much more comfortable with autonomous agents when they can see the "Thought Process" (logs) scrolling by. Visibility is a feature, not just a debug tool.
Prompt Engineering is Security: In Agentic Finance, the prompt isn't just text; it's the firewall.
Stablecoins are Critical: Using MNEE was essential because agents need stability. An agent cannot calculate ROI effectively if the currency itself fluctuates by 10% in an hour.

What's next for PRAETOR_AGI: Autonomous Treasury Sentinel

Multi-Agent Swarms: Expanding from one agent to a "Board of Directors" (one Risk Officer, one CEO, one Trader) that must vote before a transaction is signed.
Live Mainnet Deployment: Moving from simulation to a fully funded treasury on the MNEE Mainnet.
DAO Integration: Allowing a community to vote on the "Spend Limit" variable via a smart contract, making PRAETOR a truly decentralized employee.