Can we use System One models for DeFi risk management? 🤖 Some of the recent Jev demos inspired me to point one at a workflow we use at @twynexyz. When we want to understand the protocol's risk under different market conditions and parameters, we run an agent-based model. We let a bunch of simulated borrowers and credit LPs loose in the system and watch what breaks, and when. Seeing where and how it fails helps us make better calls on the core mechanics and on parameters like interest rates and liquidation thresholds. Traditionally, these agents run on simple heuristics or hand-coded probability distributions: a fixed price expectation, some sensitivity to rates, a static risk appetite, a timer for how often they check their portfolio. It's cheap to compute and gives us some signal, but it's far from how people actually behave. Real users aren't rational robots. They get scared, they get greedy, they copy their peers, they overreact to news and change strategy halfway through. Encoding that with explicit math gets messy fast. So what if we described user archetypes in plain language instead, like "risk-averse borrower who watches sentiment" or "high-leverage borrower who reacts to macro news", and let a System One model make the calls? Agents could act on much richer context: what their peers are doing, their broader goals, even market-moving headlines. That could surface emergent behavior and feedback loops that are hard to encode by hand but could break the system in the real world. For this first experiment I used @fastinoAI's GLiNER2.5-Decide, running locally on my MacBook Pro. It isn't trained for this kind of task, but it was good enough for a proof of concept. In the video, three agents start with identical ETH vaults and ride the same synthetic rally and reversal. Same model, only the persona prompt differs.