Auto-building agents plugin: Jev classifier in a computer-use cache
The author shipped a ChatGPT plugin that distills long-horizon agents into deterministic code and added a Jev-based classifier to an open-source computer-use cache to cut repeated-task cost to near zero.
Our Auto-building agents ChatGPT plugin was just published! http://chatgpt.com/plugins/plugin_asdk_app_6ac2633c40008191a028cd2643815cce This has been one of my most useful secret tools the past few months. Long-horizon agents often fail because frontier models hallucinate at 1-5%, so over 500 steps that degrades into 100% failures. The plugin distills long-horizon agents into deterministic code in a loop, so you can ask codex to build super complex reliable agents that costs up to 100x less. Often I ask Codex to build an "impossible" agent, then come back hours later or wake up to it actually working. This is runs on top of our open source computer-use cache, which reduces agentic repeated tasks to ~$0 by caching the code. I implemented a JEV based classifier into the cache last month: https://github.com/rohanarun/computer-use-cache 3 days after I open sourced it, Poetic raised $50M from Kleiner Perkins and OpenAI on basically the same idea. I found the Google researcher who publish a similar paper a month before we open sourced, and now we're working on a new geometric approach to self-attention.