AI Agent Design Patterns: ReAct, CodeAct & More | Stanford CS329Z Lecture 6 with Michael Ryan

AI Agent Design Patterns: ReAct, CodeAct & More | Stanford CS329Z Lecture 6 with Michael Ryan

Title history
  • AI Agent Design Patterns: ReAct, CodeAct & More | Stanford CS329Z Lecture 6 with Michael Ryan — observed on 20261114-2108 UTC (YouTube)
🎙 Michael Ryan 👥 1.3M 📅 November 14, 2026 ⏱ 76 min 👁 69 📄 tutorial 🧭 2026-11-14
Available in: English (current) Français

Keywords

agent designworkflowReActcontext managementorchestrator-workers

Summary

This Stanford CS329Z lecture, taught by Michael Ryan, provides a comprehensive overview of design patterns for engineering AI agents. The session begins with a concise review of the distinction between workflows (where control flow lies in code) and agents (where control flow is determined by the LLM). The speaker then introduces Anthropic’s five workflow patterns: prompt chaining, routing, parallelization, orchestrator-workers, and evaluator-optimizer, each illustrated with practical examples and trade-offs. A detailed case study of the Storm system demonstrates how these patterns can be composed to build a complex research agent that generates Wikipedia-style articles. The lecture also covers context management techniques (such as event logs, compaction, and observation masking) and reasoning-acting patterns including ReAct, Reflexion, CodeAct, and ReWOO, ending with a look at the Pi harness as a production agent case study. Throughout, emphasis is placed on the need to adapt patterns to specific contexts rather than applying them rigidly, supported by an interactive Q&A session addressing questions on modularity, benchmarking, and loop termination.

164 words

Critical Evaluation

Value of the Information & Strength of the Argument

The lecture succeeds in providing a well-structured and practically relevant framework for agent design. It clearly explains the rationale behind each workflow pattern, including latency-accuracy trade-offs, cost considerations, and criteria for when each pattern is most appropriate. The use of the Storm system as a composed example effectively illustrates how multiple patterns can integrate in a real application. The argumentation is sound, with references to Anthropic’s taxonomy and academic papers, and the speaker openly acknowledges the need for customization and the ‘half-baked’ nature of patterns. The inclusion of a live Q&A enhances the value by addressing audience concerns about scalability and benchmarking, showing critical engagement with the material.

Scientific Rigor, Source Quality, Title Accuracy

The content is based on credible sources: Anthropic’s ‘Building Effective Agents’ post and research papers such as the Storm system. As a lecture in Stanford’s official CS329Z course, it maintains high academic rigor. The speaker, a CS PhD student, demonstrates deep familiarity with the subject. The title accurately reflects the lecture’s scope, covering both workflow and agent design patterns. The presentation is organized, with clear learning objectives and a logical progression from foundational concepts to advanced patterns. The description provides direct links to the course and related resources, further supporting the video’s credibility.

216 words

Title / Content Match

Title accurately reflects content: the lecture covers AI agent design patterns including ReAct, CodeAct, and workflow patterns, matching the engineering focus described.

Quality & Reliability

9/10

Lecture delivered by a Stanford CS PhD student within an official graduate course (CS329Z). Content is grounded in established design patterns and research, with clear structure and Q&A. Minor caveat: single source (lecture) and some patterns are generic.

Key Moments

Cited Sources

Contribution & Novelties

The lecture provides a structured taxonomy of agent design patterns, synthesizing Anthropic’s workflow patterns with reasoning-acting patterns like ReAct and context management techniques. It offers practical guidance on when to use each pattern and emphasizes the need to adapt them to specific use cases. The inclusion of a real-world example (Storm) illustrates how these patterns can be composed, making the content directly applicable.

Pour aller plus loin :

132 words

Radar Profile

The profile shows a well-balanced lecture with high quantity and quality of information, a solid technical level, and strong overall reliability. The only slight dip is in technical depth (8), reflecting its nature as a broad overview rather than a deep dive into a single topic.

Reliability 9/10