MultiplAIers

Agent Guardian: AI agents that decide like you and are checked before they act. Live demo: https://amar-nuc-auto--decision-twin-web.us-east.modal.run Code: https://github.com/amarsingh01/Meta-AI-decision-twin Every real project has dozens of open threads, all waiting on you for decisions. Interruptions and decision fatigue make you the bottleneck, and agents with real access can drift off task. With Agent Guardian you write your principles once: weighted values plus hard rules. Decision twins: when an OpenClaw builder agent reaches a choice, the owner’s twin scores every option 0–10 on the owner’s principles, knocks out options that break a deal-breaker, turns the scores into probabilities and answers with its reasons. Close calls are decided and flagged for later review, so the owner isn’t interrupted. Guardian: it holds the keys; the agents have none. Every action is checked before it runs. Hard rules block secrets, uploads and unknown destinations. A Meta Llama 3.2 3B model, running locally on CPU with no API, judges the rest. Demo: Three owners decide the same question live, with the maths shown step by step. Three agents built the same site in parallel and produced three different sites. They made 15 decisions, and the owner was interrupted zero times. A replay of a reported exfiltration attempt: all 4 steps blocked. Test set: rules alone stopped 10 of 15 harmful actions; rules plus the local model stopped 15 of 15 and allowed 19 of 20 normal actions. Up next: LoRA fine-tuning of the local model on your own past decisions. Built with Meta Muse Spark, Llama 3.2, OpenClaw and Modal. Individual entry.