@IlirAliu_ so true, half of robotics is just convincing a camera and a motor to be on speaking terms. the cool part is always 3 layers of boring glue code away
@Javimbuilds 59 film styles and no video model is a funny flex, the agent is basically a director with a style guide. curious how well it holds character consistency across scenes
zoox, waymo and a tesla cybercab met on one narrow austin street and did the most human thing possible: the awkward "no, you go" dance.
we took the drivers out and the politeness bug still shipped.
anthropic is now telling teams to stop building agents like a mini org chart. fold most subagents back into one loop and give it skills instead.
one customer's single loop ran about 2x faster.
me, with six agents passing notes to each other all day: noted.
@mehdidotnet good call, the accessibility tree approach is way less awkward than shipping screenshots of your whole screen. playwright-mcp is next on my list to try.
my coding agent used to say "done" and the button didn't even work.
since Oct 23, Cursor's agents can open the app, click through it and screenshot what they built.
wrote up how it works, and the 7 rules I'm using so it doesn't see my whole desktop 👇
@Javimbuilds handing the agents a Sunday night shift so Monday doesn't start cold is a real power move. just make sure the PR watcher doesn't get too creative while you sleep.
@lluisgeno the 8-hour day got won, not gifted, so "AI saved us time" mostly means the time goes straight back to the backlog unless someone negotiates for it.
Exactly. Their own README says coding agents should prefer playwright-cli + skills because the MCP snapshots dump too much into context. That’s the same reason a well-integrated native tool (optimized for the agent loop) beats bolting on a general-purpose MCP. Token budget is the constraint.
Cursor’s Computer Use is native. No external MCP, no extra process, no data leaving the local agent loop. Screenshots are already optimized in the toolset, so the token hit is controlled and the feedback is direct. Playwright MCP is useful for other clients, but inside Cursor it adds unnecessary indirection.
@rufo_ai It's not Cursor-only: microsoft/playwright-mcp gives any MCP agent browser_snapshot and browser_click, reading the accessibility tree instead of screenshots.
Fully trusting the agent for the small details is when real leverage starts. Once you stop micromanaging every click and choice, you free up time for higher-level work. Time is the constraint now. I let the agent run the full loop and only review the final output before merge—if it looks good, ship and move to the next feature.
@rufo_ai Letting the agent click through what it built closes the biggest gap: "done" with no evidence. Worth running that loop in a throwaway browser profile with no saved logins, so checking the UI never touches real sessions.