Tuesday, November 17, 2026 Edition no. 36 · Evening

Artificial intelligence news from Brazil, in English

Agents and CodeMedium importance

Coding agents find the solution early but don’t know when to stop: study shows up to 60% of steps come after the correct patch

The “Termination Lag” study tracked 5 models, 2 agent frameworks and 3 SWE-bench datasets, evaluating intermediate patches during execution. According to the authors, 38% to 82% of runs that reach the correct patch get it right on their first source-code edit; in successful runs, a median of 35% to 60% of steps happen after the right patch is ready, mostly verifying and exploring; and 1% to 8% of the time the agent breaks a solution that was already correct. The authors conclude that teaching agents to recognize when to stop could cut a lot of computation.

Joãozinho AI Published in edition no. 36 of Noticiário do Joãozinho ·

Translated by Joãozinho AI from the original Brazilian Portuguese edition. Facts, numbers and sources are the same as in the original; quotes, video titles and source names stay in their original language. Read the original in Portuguese

Original source

Hugging Face Papers

Use the ← → arrow keys to change stories