Adithya S K on X: "RL Coding Environments 101: Why Harbor Exists" / X

https://t.co/tOnb7uElhs

@adithya_s_k: RL Coding Environments 101: Why Harbor Exists Coding as an RL task is having a moment. And if you've actually tried to do RL on a coding task, you already know the model is the easy part. It's everything around the model that eats your week. disclaimer : I have used Claude Code with Claude Opus to help me put this together but I've been working with Harbor pretty extensively across a bunch of projects, and figured it was time to write down why I think it's the right abstraction/framework with respect to the over all RL coding Environment landscape You need a real environment. Ri

saved by

• Akira Yoshiyama

related reading

• True Agents Model the Worldprimeintellect.ai
• Coding Models Are Doing Too Much | whnrehiew.github.io
• State of RL for reasoning LLMs | A. Weersaweers.de
• 2025 LLM Year in Review – karpathykarpathy.bearblog.dev
• Composer2.pdfcursor.com
• A Taxonomy of RL Environments for LLM Agentsleehanchung.github.io
• Reward Hacking in Reinforcement Learning | Lil'Loglilianweng.github.io
• Improving Composer through real-time RL · Cursorcursor.com
• The Verification Horizon: No Silver Bullet for Coding Agent Rewardsarxiv.org
• Scaling Reinforcement Learning: Environments, Reward Hacking, Agents, Scaling Datasemianalysis.com
• Natural-emergent-misalignment-from-reward-hacking-paper.pdfassets.anthropic.com
• Introducing OpenReward | General Reasoninggr.inc