browser-use/browser-harness

Browser Harness wires up a real browser to an LLM task loop and, critically, adds a self-healing layer so that when a button moves, a class name changes, or a page flow shifts, the agent recovers rather than crashing. That last part is the whole product. Anyone can glue Playwright to an LLM and call it agentic automation. The painful part is the three AM failure where the target site updated a checkout flow and your whole pipeline is dead. This project specifically targets that failure mode, attempting to re-derive element targets at runtime instead of relying on brittle selectors baked at record time. It is open-source and the repo is early but active. Honest reservation: self-healing in browser agents is genuinely hard and most implementations handle the easy cases only. Worth running against your actual use case before committing. -> Best for: AI engineer or solo founder building LLM-driven automation over third-party web UIs