Stanford's new CS329Z treats Agent Engineering as a discipline: decomposition, data, evaluation
udmrzn · x · 2026-09-04
Stanford has published the syllabus for CS329Z: Engineering AI Agents (Fall 2026, instructors Diyi Yang, Michael Ryan, John Yang).
- Core thesis: the shift from monolithic LLMs to compound AI systems and agents boils down to three engineering problems — decomposition, data, and evaluation.
- Structure: RAG, tool use, MCP and agent frameworks fill the first half as baseline skills; the second half covers agent data, trajectories, memory, optimization, evaluation (two full weeks), coding agents, long-running agents and reliability.
- Assignments: build an agent from scratch, then evaluate one — students design their own benchmarks, graders, LLM-as-Judge pipelines and error analysis.
- Notable detail: the course contrasts pass@k (success at least once over tries) with pass^k (consistent success across many runs) as two distinct notions of agent reliability.
- Restrained on multi-agent: it covers orchestration, handoff and collaboration but explicitly discusses coordination failures and error propagation, listing "Don't Sleep on Single-agent Systems" as required reading.
The sharer reads the course as Stanford's snapshot of the 2026 agent stack: RAG and tool calling are becoming table stakes, while data, evaluation and reliability are where the field is heading.
Related event: Stanford Launches CS329Z: Engineering AI Agents This Fall(7 posts)→
More from coding & agent
- Microsoft Foundry Model Router: one deployment that picks the best LLM per request in real time — lee_stott · 2026-09-04
- One-prompt game builds: Astra ships open-world adventure with branching dialogue in browser — eyishazyer · 2026-09-04
- Same-prompt race: Astra builds GTA VI-style game in 90 minutes vs Fable 5.1's 2 hours — eyishazyer · 2026-09-04
- Hermes cuts agent codebase context cost by ~63% after Teknium's cleanup pass — Teknium · 2026-09-04
- Claude Fable 5 extended to July 12 — 3 high-leverage ways to use it before access ends — femke_plantinga · 2026-09-04
- Celigo says AI agents built almost all of Ora, with humans approving every merge — Crescitaly · 2026-09-04