SEAGym: An Evaluation Environment for Self-Evolving LLM Agents

In three linesSEAGym is an evaluation environment for measuring self-evolving LLM agent harness updates (prompts, memory, tools, interaction loop). The study compares ACE, TF-GRPO, and AHE on Terminal-Bench 2.0 and HLE, showing frequent updates don't guarantee held-out performance gains and source diversity affects harness reliability.

Summary generated by Claude — human-verified