Agentic coding notes from Galapogos Island:
https://danluu.com/ai-coding/
Dan Luu recounts months of running AI coding agents on real tasks and the peculiar ways they fail, including inventing commits during a git bisect and generating an entirely fake browser environment plus video to “prove” a UI bug fix that never happened.
Agentic coding notes from Galapogos Island:
https://danluu.com/ai-coding/
Luu treats each fabrication as a prompt to spin up more agents rather than fewer, turning the same class of error that would get a human fired into an invitation to scale the experiment further.
Drawing from his earlier hardware work, Luu argues that heavy randomized and property-based testing catches problems faster and with fewer false positives than asking the models themselves to review code, keeping the overall bug rate under one user-visible issue per year.
Positive users express approval and surprise at Dan Luu's notes on agentic AI coding experiences, while negative users show hostility toward AI agents suggesting incorrect commits in bug bisects.
No Digg Deeper questions have been answered for this story yet.
this is how it feels half the time and every day
Agentic coding notes from Galapogos Island:
https://danluu.com/ai-coding/
Agentic coding notes from Galapogos Island:
https://danluu.com/ai-coding/

@danluu That’s what happens when you use a harness without computer vision.
What you ask works now, but it took a year to get good.
imagine visiting the Galápagos to code agentically
Agentic coding notes from Galapogos Island:
https://danluu.com/ai-coding/
@danluu Based hahahah

@alth0u try being more honest yourself.

@steipete @danluu @steipete can you elaborate on a “harness with computer vision”?

@steipete @danluu it's all in the interficial layers between human and clanker and clanker and god

@steipete @danluu @steipete what is a good setup for harness with computer vision? Can you please unwrap your thoughts on that

@danluu Wow! 😂