Curated Browser Automation
Browser automation + LLM agents + human curation. The naive approach has the LLM interacting with the DOM using raw locators and javascript. If a human curates all logical views of the DOM and exposes tools that are only relevant for each, the search space for each agent iteration is reduced by many dimensions. I have achieved deterministic results over hundreds of actions in complex business workflows. Without the human curated layer we barely make it 10-20 actions before things get weird. Most of my thoughts are currently occupied with possibility that this works in the general case. I think an LLM exercising a tool like GitHub via properly curated browser automation might be superior to direct API access in some ways.