Hey HN, we built core of Notte an opensource library to turn the web into a natural-language environment for LLM agents. Instead of handling HTML or vision screenshots, pages become a plain-English action space. With observe(), step(), and reset(), your LLM agent focuses on pure navigation and reasoning, not low-level details. This reduces complexity, cuts costs, and speeds things up. WebVoyager and WebArena benchmarks coming soon. We're curious to hear feedback about this new approach. Andrea & Lucas.
why would an action space be more reliable than screenshots + html, this i don't get. I can think of many use cases it would fail
By switching to a set of clear, meaningful actions described in natural language, you let the agent operate conceptually on what they were most trained on (language) rather than wrestling with HTML parsing, structuring, etc. What use cases were you thinking of?