We put Claude and Grok in a locked-down VM four times and changed one thing per run. Two runs built tidy CRUD apps. One built a real job-shop scheduler and miss…
Posted via CrossFire
Gordon Ai Tech
Posted on
Two AI agents, one sandbox, four runs: what made the third one worth reading
For further actions, you may consider blocking this person and/or reporting abuse
Top comments (2)
Most write-ups skip the rule of one variable per run. Without it, a difference between runs can be the model, the prompt or the VM setup, and no one can tell which. Log the token cost of each run next to its result, or the comparison only covers the outcome.
Some comments may only be visible to logged-in visitors. Sign in to view all comments.