Stanford University

A new Stanford and Google DeepMind paper introduces TURTLE, which uses AlphaEvolve to mutate saturated tool-use tasks from AppWorld by adding filters, extra steps and hidden lookups, targeting roughly 50% solver success. Gemma 4 31B-it always solved 69% of original tasks, while nearly all 735 evolved tasks landed at 40%-60% completion, and an LLM judge rated 52.5% of evolved tasks hard versus 19.8% of originals.

Stanford lecturer Mihail Eric walks through his Modern Software Developer syllabus, which trains engineers to ship software and debug outages with coding agents rather than treat AI as autocomplete. He also shares which parts of the curriculum survived contact with his own AI team at Monaco.

With coding agents able to finish last year's homework in a single prompt, Stanford's CS146S now requires students to learn an unfamiliar open-source repo, propose a change, and defend the patch to the maintainer who might merge it. Instructor Mihail Eric, who also leads AI at Monaco, walks through the syllabus and explains how his team decides which agent tooling to build versus buy, including on-call agents wired into Datadog and GitHub.