96.8% of tasks done perfectly. No other tool tops 84%.
Playwright MCP re-reads the page through a tool call at every step, the browser sits idle while the model thinks, and every extra round trip is another chance for a long task to fall apart. On those 31 jobs, agent-driven Playwright ends 29% of tasks short of a perfect score.
In ego lite, every page reaches your agent as a Snapshot: a compressed accessibility tree with stable @N refs it can act on instantly, several actions per JavaScript turn. The agent spends its time doing the task, not parsing pages: 96.8% of the same 31 tasks end perfect, at the fastest average task time of the five tools measured.

