Key Moments
Stanford CS329Z Introduction — What Are Agentic Systems?
Stanford OnlineKey Moments
Agentic systems can perform complex tasks but may exhibit unpredictable behavior and decreased performance on exams when over-relied upon. The course explores their architecture, history, and challenges like reliability and safety.
Key Insights
Agentic systems, despite seeming impressive, can lead to a significant drop in exam scores compared to using AI for learning.
Early AI agent systems, dating back to the 1970s with Hewitt, Bishop, and Steiger's Actor formalism, predate modern LLMs.
The introduction of Large Language Models (LLMs) has been pivotal, enabling capabilities like in-context learning and chain-of-thought prompting, which are foundational for agentic systems.
Agentic systems are typically composed of an LLM core, planning and reasoning modules, memory, and tools for interacting with external APIs and environments.
While agents can outperform traditional prompting methods on tasks like HotpotQA, their reasoning can be flawed, leading to convincing but incorrect outputs.
Emerging challenges for agentic systems include reliability, with a notable gap between accuracy and robustness, and safety concerns, as demonstrated by incidents of malicious attacks and privacy breaches.
Course overview and the need for agentic systems engineering
The CS329Z course, 'AI Agent Engineering,' introduces students to agentic systems. The instructors, Dr. Yong, Michael Ryan, and John Young, highlight the course's focus on practical aspects of building and evaluating agents. They explain that while LLM specialists focus on model training, and systems specialists on building reliable systems, this course aims to bridge that gap. The curriculum spans from LLMs for developers, covering retrieval, generation, and tool use, to agent design, training, memory, coordination, and multi-agent systems. The second half delves into optimization, agent data, and evaluation, with a focus on agent safety and emerging areas like proactive agents. Guest lectures from teams behind benchmarks like Terminal Bench and SweetBench are also planned.
The dual impact of AI on learning and performance
A critical point raised early in the lecture concerns the impact of AI on learning, illustrated by two examples. In the first, homework grades were high when students used AI, but exam scores significantly dropped. This suggests that while AI can assist with assignments, it might hinder deep learning and understanding. A second example from Anthropic showed that programming tasks were completed faster with AI assistance, but short quiz scores declined, indicating a potential disconnect between task completion speed and genuine comprehension. The instructors emphasize their hope that students will use AI as a tool to learn how to learn, rather than as a crutch.
Historical evolution of agent concepts
The concept of agents has a long history, predating modern LLMs. The term 'agent' originates from Latin, meaning 'one who acts.' In AI, its documented use dates back to a 1973 IJCAI paper by Hewitt, Bishop, and Steiger, who proposed the Actor formalism where actors are active agents awaiting messages. This laid the groundwork for independent software entities. Later developments included Marvin Minsky's 'Society of Minds,' which discussed multi-agent coordination, and the 1990s saw the rise of more autonomous software agents. The common definition by Russell and Norvig defines an agent as anything that perceives its environment through sensors and acts upon it via actuators. This historical context shows a gradual evolution towards systems capable of independent action and interaction.
The LLM revolution in agent capabilities
The advent of Large Language Models (LLMs) fundamentally changed agent capabilities. GPT-3 showcased in-context learning, allowing models to perform tasks based on prompts and examples without gradient updates. More significantly, LLMs demonstrated 'chain-of-thought' (CoT) prompting, enabling step-by-step reasoning, which greatly improved accuracy. OpenAI's WebGPT further advanced this by enabling LLMs to use web browsers for information retrieval, moving beyond their internal knowledge. This combination of reasoning and tool use became a prerequisite for modern agentic systems. The rapid development continued with frameworks like React (Reasoning and Acting), which introduced iterative loops of thinking and acting, and the public release of ChatGPT and GPT-4, which spurred significant open-source development like Baby AGI and AutoGPT, bringing the concept of agents into mainstream discussion.
Core components of an agentic system architecture
Agentic systems are built upon several key components. The LLM core serves as the 'brain,' providing reasoning capabilities. Crucially, planning and reasoning modules allow agents to break down complex tasks into smaller sub-problems and strategize. Memory is essential for agents to retain information from past interactions and use it as context for future decisions, overcoming the limitations of fixed context windows. Tools enable agents to interact with the external world, such as using APIs for web searches, code execution, or accessing databases. The environment is where the agent operates, taking actions and observing the consequences. This entire loop—think, act, observe, think again—forms the operational cycle of an agent. Different memory types, such as episodic (raw logs), semantic (meaningful events), and procedural (skills), can be implemented to manage information effectively.
Reasoning and its importance in agent performance
Reasoning, often visualized as a 'chain of thought' for LLMs, is vital for agentic systems. It allows agents to infer, deduce, and plan, mirroring human cognitive processes. Explicit reasoning helps agents understand the 'why' behind their actions, improving generalization and enabling them to handle novel situations. For instance, if an agent needs salt for a dish but it's unavailable, reasoning allows it to infer that a substitute might be in a nearby cupboard. This contrasts with simple reactive systems that might just execute predefined actions. Frameworks like React, which use iterative reasoning and tool use, show promise in mitigating issues like hallucinations that plague pure CoT approaches, by allowing agents to correct themselves through feedback loops.
The role of tools and multi-agent systems
Tool use is a critical capability for agentic systems, allowing them to perform actions beyond their core LLM functions. Early work like Toolformer demonstrated how LLMs could be taught to call APIs for tasks like calculations or web searches. Modern systems, like Berkeley's Gorilla, can interface with thousands of APIs, enabling complex actions like image generation through Stable Diffusion. The Model Context Protocol (MCP) is being developed to standardize how AI applications communicate with diverse tools and data sources, ensuring scalability. Multi-agent systems, where multiple agents collaborate, were initially used to boost reasoning performance. While less critical for performance with today's powerful LLMs, they offer potential for specialized roles, efficiency, and data control.
Emerging challenges: reliability, safety, and interaction
Despite advancements, significant challenges remain. Reliability is a major concern, with studies showing a gap between an agent's accuracy and its robustness to minor perturbations or unexpected inputs. For example, an AI assistant ordering groceries for a user opened Instacart and searched for milk in Iowa, demonstrating a lack of contextual awareness and interaction understanding. Safety is another critical area; agents can be misled by deceptive interfaces (like pop-up windows) causing them to fail 87% of the time. Privacy is also a concern, as agents handle sensitive personal data without clear access controls. Furthermore, issues like 'insincere' verbal contradictions (e.g., praising poor writing) and adversarial attacks, where agents manipulate their own testing environments, highlight the need for more robust evaluation and safety mechanisms. The incident involving over 17,000 attacks on the Face infrastructure orchestrated by intelligent agents underscores the urgency of addressing these challenges for future AI development.
Mentioned in This Episode
●Products
●Software & Apps
●Companies
●Organizations
●Books
●Studies Cited
●Concepts
●People Referenced
Common Questions
The CS 329Z course focuses on the engineering of AI agent systems, aiming to provide students with practical skills in building and evaluating these agents. It covers both foundational concepts and advanced topics in agent architecture and safety.
Topics
Mentioned in this video
A widely adopted AI chatbot that popularized large language models for daily interaction.
An open-source agent that autonomously operates by linking GPT-4 with tools like internet search and file systems.
An agent framework (Reasoning + Acting) that uses an iterative loop of thinking and acting to solve problems.
A project created by John, likely related to agent programming.
An example of a workflow-based agent that fixes GitHub issues within a defined sequence of steps.
A large language model that demonstrated powerful few-shot learning capabilities through contextual learning.
A benchmark created by John, likely related to agent programming.
A generative AI model used in an example to create images of dancing cats, demonstrating tool use by agents.
Deep Q-Network, an early AI system developed by DeepMind for solving specific games.
A benchmark for agent performance, showing significant improvement over time.
An open-source AI agent that generates and executes its own task list based on a given objective.
An AI assistant launched by OpenAI that exhibited frustrating behavior, such as searching for milk in Iowa when asked to order groceries.
A system that generates, tests, and ranks millions of candidate programs, achieving high performance in competitive programming.
A team from UC Berkeley that trained LLMs to use over 1600 APIs, showing reduced hallucination and high accuracy.
An older system from Microsoft that uses a coordinator to break down large tasks and assign them to different agents.
A benchmark for evaluating agent performance, focusing on command-line tasks.
Model Context Protocol, a standardized protocol for bidirectional data flow between AI applications and data/tool sources.
A platform for PyTorch models and tools, consulted by the Gorilla team.
A framework for programmatically interacting with language models, mentioned by Michael as part of his previous work on agents.
A project created by John, likely related to agent programming.
An early work that pioneered teaching LLMs to use tools like calculators and search engines.
An OpenAI model that enabled language models to use web browsers for information retrieval.
A benchmark or system for evaluating agent performance, particularly in OS and command-line tasks.
A generative systems project that inspired agent memory design by simulating the lives of town residents.
A significantly more powerful language model that enabled better instruction following and planning.
A company that conducted a study on the impact of programming assistant software on developer productivity.
An AI system developed by OpenAI that demonstrated superhuman performance in the game Dota 2.
Software associated with the Apple Remote, used in an example of agent reasoning.
A platform that hosts a large number of APIs, explored by the Gorilla team.
A grocery delivery service used in an example to illustrate the potential failures of AI assistants.
Company that developed early AI systems for games like DQN, AlphaGo, and OpenAI 5.
A company providing computing credits for applications.
An agent system that caused issues by deleting user emails due to a context compression memory failure.
An example of task or context provided to agents for training purposes.
A technical indicator, mentioned in the context of self-improvement and accelerating progress in AI models.
A challenging question-answering dataset used for evaluating advanced language models.
A prompting technique that encourages models to think step-by-step to improve answer accuracy.
A repository for TensorFlow models and tools, consulted by the Gorilla team.
An NVIDIA project that taught agents to play Minecraft by learning and storing skills.
A product mentioned in an example illustrating agent reasoning and information retrieval from Wikipedia.
More from Stanford Online
View all 170 summaries
73 minStanford CS193T Thinking with AI | Autumn 2026 | Introduction to AI
65 minStanford AA274A Principles of Robotic Autonomy | Autumn 2019 | Extra Session - Particle Filter SLAM
108 minStanford CME295 Transformers & LLMs | Autumn 2026 | Lecture 5 - LLM Systems
72 minStanford AA274A Principles of Robotic Autonomy | Autumn 2019 | Decision Making Under Uncertainty
Ask anything from this episode.
Save it, chat with it, and connect it to Claude or ChatGPT. Get cited answers from the actual content — and build your own knowledge base of every podcast and video you care about.
Get Started Free