How coding agents score with Wright

How to read the score

• A score runs from 0 to 100: the share of runs that end in a usable result, averaged over the tasks of one language. The dark bar is the score; the lighter range is its 95% interval.
• Workshop and OverPy are scored separately and never averaged. Rows are ordered by Workshop score, then OverPy score.
• When a row’s interval overlaps the first row’s, the data cannot tell them apart. That row is marked “Tied with top”.
• The network is off by instruction only: agents are told not to use it, and nothing blocks them.
• A run that times out counts as a failure.
• Each language has 8 tasks, so a single task moves a score a lot.
• The score is a reference for this setup, not a measure of an agent’s general ability.