Excited to share that two of our papers on self-improving agents have been accepted to 𝐄𝐌𝐍𝐋𝐏 𝟐𝟎𝟐𝟔! Code for both is released in A-Evolve. 1. 𝐇𝐚𝐫𝐧𝐞𝐬𝐬 𝐔𝐩𝐝𝐚𝐭𝐢𝐧𝐠 𝐈𝐬 𝐍𝐨𝐭…

• Achieves top-tier benchmark results with 3 lines of code and zero manual harness engineering • Pushes a single Claude Opus-4.6 base model to #1 on MCP-Atlas and #2 on ARC-AGI • Autonomously…

A-Evolve treats the entire agent as a mutable directory — prompts, skills, tools, memory — and runs evolutionary optimization on the filesystem instead of the weights. Each accepted mutation gets…

Self-improving agents need more than random prompt tweaks. A-Evolve is an open-source Python framework for evolving LLM agents across domains, benchmarks, and evolution algorithms. It helps you test…

We introduce Adaptive Auto-Harness: • a stateful multi-agent evolver that preserves cross-cycle knowledge; • a harness tree that routes each task to a relevant harness at solve time; • targeted human…

Harness-benefit is different: it is non-monotonic. Weak agents often fail to activate the relevant harness—or fail to follow it over long trajectories. Mid-tier agents benefit most, while strong…

🐾 A-Evolve: The Universal Infrastructure for Self-Improving Agents Elevate any AI agent to state-of-the-art performance with just 3 lines of code! Zero manual setup needed. Dive in and see your…