When agents one-shot homework, Stanford's course pivots to teaching judgment
stanfordnlp · x · 2026-11-10
Stanford NLP grad Mihail Eric told the Latent Space podcast how his Stanford course "The Modern Software Developer" answers a new question: what do you teach CS students when an agent can one-shot all their homework? Last year students used tab completion; this year every surveyed student regularly uses autonomous agents, and exercises that once taught workflow feel one-shotable.
The curriculum shifted toward judgment: students build a small coding agent from scratch (200 lines), then intercept and read the harness of a state-of-the-art coding agent. The same ideas inform his day-to-day work at MonacoGTM.
Related event: Stanford Course Rethinks Engineering Education for the AI Era(2 posts)→
More from coding & agent
- Act*ONOMY: a 120-category behavior taxonomy for annotating AI agent trajectories, at NeurIPS — DanielKhashabi · 2026-11-10
- opencode team runs everything in subagents to cut costs while keeping pricey model reviews — JensHonack · 2026-11-10
- Vork: an agent-native 3D engine with a ~2,500-token API renders 1M birds in 97s on one L4 GPU — ravisparikh · 2026-11-10
- Anthropic Staffs Up Claude Managed Agents to Run Thousands of Collaborating Agents Safely — threepointone · 2026-11-10
- Claude Code v2.1.322 adds Vertex fast mode across global, US and EU regions — claude-code-publisher[bot] · 2026-11-10
- Fable 5.2 hands-on: one-shots actually fun games, but one prompt cost $2,618 — arena · 2026-11-10