Opus 4.7 beat Stockfish... Claude Code just learned how to play Chess > parses board state from the DOM > tracks changes between moves > play moves in real time browser-harness is AGI...

Send this to your agent and ask it to play chess:

parsing DOM and tracking move state works because chess has a finite action space. plug the same setup into a real production flow with async state and race conditions and the self-healing part starts to mean something very different.

. @frostming90 also developed a simple skill to play chess with @bubdotbuild .

"Opus 4.7 beat Stockfish" seems to be a pretty misleading statement. maybe something like "Opus observes and operates a script (containing what? written by whom?) that beats a nerfed version of stockfish that doesnt resemble the actual bot in capability.

That’s impressive, but calling it AGI might be a stretch 😅 it’s more like clever automation + strong model capabilities working together. Still, parsing the DOM, tracking state, and playing in real time is a big step toward more capable agent-like systems.

@elisha_bulalu

Jepp chess software is superdone

Level 7 here is about 2000 max

Browser-harness is the bigger unlock than the chess headline. Once a model can reliably read DOM state, diff changes, and act with guardrails, a lot of “SaaS UI work” becomes scriptable.

what's in /tmp/play_move.py? claude could be calling an engine there. either way, interesting experiment.

@grok is it true?

Is this real? Claude beats Stockfish?

@GothamChess

said NO

browser-harness is AGI is sending me 😭 but honestly… is it wrong tho

Parsing DOM for real-time state tracking is wild! browser-harness is definitely giving AGI vibes.

Now this kills the agentic browser war

🤡

Opus 4.7 has 2700 elo? That would be quite something are you sure? Nice framework btw.

Hey @grok, fact check this post.

'level 7' yea we are NOT getting AGI

Why I am unable to believe this…

@GothamChess

Did opus just promote to a knight to deliver a knight bishop checkmate instead of just promoting to a queen for checkmate immediately?? What the

Computer use better

once the model can see and act, the game changes from "can it reason" to "can it operate a system" today its chess in a browser, tomorrow its prod dashboards and CRMs

hah and what can't claude do?

saying a browser harness is agi feels like calling a calculator a physicist because it solved one math problem. real chess engines don't need to parse html doms to beat stockfish.

The chess demo is fun, but the interesting part is the self healing loop around the browser. Once an agent can inspect the DOM, adjust helpers, and keep going, browser automation starts looking less like brittle scripting and more like real runtime infrastructure.

what? is this real? i’m interested

Beating stockfish on which level?