AgentEvolver

Why we included this project

Most agent training pipelines plateau once the fine-tuning run ends, and if you have hit that wall, AgentEvolver is worth a look. It is a training framework that lets an agent keep improving after training by generating its own tasks, navigating its own learning path, and attributing which intermediate steps actually produced a good outcome. That cuts down on hand-built datasets and manual rollout collection, which matters if your budget for curating agent data is small. The repo also includes a service-oriented architecture with environment sandboxes, plus a game arena for multi-agent social reasoning in Avalon and Diplomacy that doubles as a training and evaluation ground. Benchmarks on AppWorld and BFCL-v3 show gains over base models at the same parameter count, so if you are weighing whether to build a self-evolving training pipeline from scratch, this is a credible reference to start from.