Long-Horizon Agents Need Experiments, Not Just Prompts — Erina Karati

Multi-agent AI systems often struggle with maintaining social consistency and accurate state management over long durations. Project Paradox addresses this by providing a modular framework where agents track memories, emotions, and trust scores. To overcome issues like source attribution loss and the degradation of rumors into facts, an autoresearch layer acts as a meta-system that runs controlled experiments. By defining specific scenarios and using a balanced scorecard to measure metrics like reach, provenance, and uncertainty preservation, developers can iteratively refine the agent protocol. This approach replaces subjective prompt tuning with systematic, data-driven improvements, ensuring that agents maintain context and logical behavior. Ultimately, building robust long-horizon agents requires treating the system as a lab bench, where small, tested policy changes are applied to a constrained surface to improve overall societal behavior within the simulation.