AgentFlow introduces a typed graph DSL and feedback-driven optimization loop for automatically synthesizing multi-agent harnesses that discover security vulnerabilities, achieving state-of-the-art results on …
This paper introduces Mimosa, an evolving multi-agent framework that automatically synthesizes and iteratively refines task-specific workflows for autonomous scientific research, achieving 43.1% success rate on …
Letta agents achieve 74% accuracy on the LoCoMo memory benchmark using simple filesystem operations instead of specialized memory tools, suggesting that agent capabilities matter more than complex retrieval mechanisms …
