AI Pioneer Shows The Power of AI AGENTS - "The Future Is Agentic"

Matthew Berman
29 Mar 202423:47

TLDRDr. Andrew Ning's talk at Sequoia emphasizes the transformative power of AI agents, highlighting their potential to outperform current models through agentic workflows. He discusses the significance of iterative processes, diverse agent collaboration, and the use of tools and reflection to enhance AI performance. Ning's insights suggest that the future of AI will be heavily influenced by the development and integration of agentic design patterns, potentially leading to significant productivity boosts and advancements towards AGI.


  • 😀 Dr. Andrew Ning highlights the significant potential of AI agents, emphasizing that they are not just large language models but tools capable of iterative improvement and specialization.
  • 🤖 He explains 'agentic workflows' where AI agents, through multiple iterations and specializations, can perform tasks more effectively than traditional models, offering the example of GPT-3.5 outperforming GPT-4 in specific tasks when used in an agentic workflow.
  • 🧠 The talk introduces the concept of AI agents working together with different roles (like writing, reviewing, and fact-checking), which mirrors collaborative human workflows and enhances the overall quality and effectiveness of the AI output.
  • 💡 Sequoia Capital is highlighted for its impressive track record in Silicon Valley, with investments contributing significantly to the NASDAQ's value.
  • 🔍 Dr. Ning elaborates on 'reflection' and 'tool use' as key AI capabilities, where AI models reflect on and improve their outputs and use external tools to enhance their functionality.
  • 📊 The benefits of multi-agent systems are discussed, showing that agents can collaborate to achieve complex tasks more efficiently than single-agent or non-agentic processes.
  • 👥 The concept of planning and multi-agent collaboration is discussed as emerging and potentially transformative, enabling AI systems to plan steps and collaborate, thus achieving tasks that are complex and require nuanced execution.
  • 🚀 The transformative potential of agentic AI systems is linked to their ability to iterate and refine their approaches, leading to significantly better outcomes over traditional AI models.
  • 🌐 Dr. Ning also touches on the practical applications of agentic AI in coding, using AI to generate, review, and debug code more effectively than traditional methods.
  • 🔧 The discussion concludes with a look forward to the continued evolution of AI agents, suggesting that as AI technology and agent-based models evolve, they will become more integral to various applications and industries.

Q & A

  • What is Dr. Andrew Ning's background and why is he considered a leading mind in AI?

    -Dr. Andrew Ning is a computer scientist, co-founder of Google Brain, former Chief Scientist of Baidu, and has been educated at UC Berkeley, MIT, and Carnegie Mellon. His extensive experience and contributions to the field make him a leading mind in artificial intelligence.

  • What does Dr. Ning mean by 'agentic workflow'?

    -An 'agentic workflow' refers to a process where multiple AI agents, each with different roles and capabilities, collaborate and iterate on a task to achieve a better outcome. This is contrasted with a non-agentic workflow where a single model generates output without iterative refinement.

  • How does the agentic workflow improve results compared to zero-shot prompting?

    -The agentic workflow improves results by allowing multiple agents to work together, each contributing their specialized functions, and iterating on the task. This collaborative and iterative process leads to more refined and higher quality outcomes than zero-shot prompting, where an AI model is asked to generate an answer without any further interaction or refinement.

  • What is the significance of the stat regarding Sequoia and the NASDAQ?

    -The stat highlights Sequoia's successful track record in investing in technological companies. It shows that 25% of the total market capitalization of companies listed on the NASDAQ are those that Sequoia has invested in, indicating their ability to identify and support high-performing tech ventures.

  • How does tool use enhance the capabilities of AI agents?

    -Tool use allows AI agents to utilize pre-existing tools and libraries that have been developed for specific tasks. By integrating these tools into the AI's workflow, the agents can perform more complex tasks and generate more accurate outputs without the need for custom coding or extensive development.

  • What are the benefits of multi-agent collaboration?

    -Multi-agent collaboration enables different AI agents, each potentially powered by different models, to work together on a task. This can lead to a diversity of perspectives and specialized contributions, resulting in more robust and effective solutions. It also allows for the simulation of human-like teamwork and problem-solving dynamics.

  • What is 'reflection' in the context of AI agents?

    -In the context of AI agents, 'reflection' is a technique where an AI model is prompted to review and improve upon its own output. This process of self-assessment and revision can lead to more accurate and refined results as the AI model iteratively enhances its initial response.

  • How does Dr. Ning envision the future of AI applications with the incorporation of agentic workflows?

    -Dr. Ning envisions a future where AI applications become more powerful and versatile through the use of agentic workflows. He expects the set of tasks that AI can perform to expand dramatically as these workflows become more prevalent, leading to significant productivity boosts and the potential for AI to take on more complex and nuanced tasks.

  • What is the importance of fast token generation in agentic workflows?

    -Fast token generation is important in agentic workflows because it allows for more iterations to occur in a shorter amount of time. This rapid back-and-forth between agents can lead to quicker refinement of tasks and improved outcomes, as the AI agents can react and adjust to each other's contributions more swiftly.

  • How does Dr. Ning suggest we adapt to the slower response times of agentic workflows?

    -Dr. Ning suggests that we need to adjust our expectations and be more patient with agentic workflows. Instead of seeking immediate responses like we do with web searches, we should be willing to allocate more time for AI agents to complete their tasks, understanding that the iterative process can lead to higher quality results, even if it takes longer.



🧠 Dr. Andrew Ng's Insights on AI Agents

This paragraph introduces Dr. Andrew Ng's talk at Sequoia, emphasizing his optimism about AI agents. Dr. Ng, a renowned computer scientist and co-founder of Google Brain, discusses the potential of agents powered by models like GPT 3.5 to achieve reasoning capabilities akin to GPT 4. The speaker expresses their own bullishness on agents and plans to provide a step-by-step walkthrough of Dr. Ng's talk, highlighting the significance of his insights. The paragraph also provides background on Dr. Ng's achievements and his platform, Corsera, which offers free learning resources in various fields including computer science and math.


🚀 Enhancing AI Performance with Agentic Workflows

In this paragraph, the speaker delves into the concept of agentic workflows and their superiority over non-agentic ones. It explains how agentic workflows mimic human iterative processes, leading to better outcomes. The speaker uses the example of writing an essay to illustrate the difference between non-agentic (one-shot) and agentic (iterative) approaches. They also discuss the power of having multiple agents with different roles and the collaborative benefits of this setup. The speaker then presents a case study showing that an agentic workflow with GPT 3.5 outperforms zero-shot GPT 4 and even approaches the performance of GPT 4 with reflection and tool use. The importance of this finding is emphasized, suggesting a significant shift in how AI applications are built.


🛠️ Design Patterns in AI Agents

This paragraph focuses on the broad design patterns observed in AI agents. The speaker categorizes the current state of agents as a chaotic but promising space with significant research and development. They outline four key design patterns: reflection, tool use, planning, and multi-agent collaboration. Reflection involves prompting the language model to improve its own output. Tool use allows the model to utilize custom-coded or existing tools for specific tasks. Planning involves giving the model the ability to plan steps, leading to more thoughtful and effective results. Multi-agent collaboration refers to the use of different agents working together, each powered by potentially different models, to achieve a common goal. The speaker provides examples and explains how these patterns can significantly enhance the productivity and performance of AI systems.


🌟 The Future of AI: Agentic Reasoning and Beyond

In the final paragraph, the speaker discusses the future of AI, particularly the expansion of tasks AI can perform through agentic workflows. They suggest that the immediacy of human-like responses from language models may become less important as we learn to耐心等待 the more thoughtful and iterative outputs from AI agents. The speaker also highlights the potential of faster token generation rates, which could make the iterative process of agents much quicker and more efficient. They express excitement about the upcoming models like GPT-5 and the advancements in inference speed, which could further enhance the capabilities of AI agents. The speaker concludes by positing that agent workflows could bring us a step closer to achieving AGI (Artificial General Intelligence), framing it as a journey rather than a destination.




AI Agents, or Artificial Intelligence Agents, refer to systems that can perform tasks autonomously, often with the ability to interact with their environment and make decisions based on data. In the context of the video, AI Agents are seen as the future of artificial intelligence, capable of reasoning and collaboration to achieve complex tasks. Dr. Andrew Ng emphasizes the potential of AI Agents to surpass the capabilities of current models like GPT 3.5 and GPT 4 through agentic workflows, which involve iterative processes and collaboration between different agents.

💡GPT 3.5 and GPT 4

GPT 3.5 and GPT 4 are iterations of the Generative Pre-trained Transformer models developed by OpenAI. These models are advanced language models that can generate human-like text based on the input they receive. In the video, Dr. Andrew Ng discusses how these models can be enhanced through the use of AI Agents, which can reason and iterate on tasks to achieve outcomes that may not be possible with the models alone. The comparison is used to illustrate the potential power of agentic workflows in improving AI performance.

💡Dr. Andrew Ng

Dr. Andrew Ng is a renowned computer scientist and educator, known for his contributions to the field of artificial intelligence and machine learning. He is the co-founder of Coursera, an online learning platform, and has held significant roles at organizations like Google Brain and Baidu Research. In the video, his expertise and insights on AI Agents and their potential impact on the future of AI are highlighted, emphasizing his bullish stance on the agentic approach to artificial intelligence.

💡Agentic Workflow

An agentic workflow refers to a process where AI Agents collaborate and iterate on tasks to achieve better outcomes. This approach mimics the way humans work, by planning, revising, and refining a task through multiple iterations. In the context of the video, Dr. Ng explains that agentic workflows allow for multiple agents with different roles and capabilities to work together, leading to improved performance and results that are more aligned with human-like reasoning and problem-solving.


Sequoia refers to Sequoia Capital, a prestigious venture capital firm known for investing in successful technology companies. In the video, it is mentioned to emphasize the significance of Dr. Ng's talk, as Sequoia is a highly respected institution in the tech industry. The firm's portfolio includes major tech companies that contribute to a large portion of the NASDAQ market capitalization, showcasing their ability to identify and invest in promising technologies and innovations.


In the context of AI and the video, reflection refers to a technique where a language model is prompted to analyze and improve upon its own output. This process involves the AI agent reviewing its generated content, identifying potential improvements, and producing a revised version. Reflection is one of the design patterns discussed by Dr. Ng that can significantly enhance the performance of AI Agents, as it allows for a more iterative and self-correcting approach to problem-solving.

💡Tool Use

Tool use in AI involves providing language models with tools or functions that they can utilize to perform tasks. These tools can be custom-coded or pre-existing functions that extend the capabilities of the AI Agent beyond its base language model. In the video, Dr. Ng discusses how tool use allows AI Agents to perform more complex tasks, such as web scraping or mathematical computations, by integrating these tools into their workflow, thus enhancing their effectiveness and versatility.

💡Multi-Agent Collaboration

Multi-Agent Collaboration refers to a system where multiple AI Agents work together to achieve a common goal. Each agent may have a different role, personality, or specialized function, and they interact and iterate to refine their collective output. In the video, Dr. Ng highlights the potential of multi-agent collaboration as a powerful design pattern, where the collaboration between agents can lead to better problem-solving and decision-making, similar to how humans work together in teams.


Planning in AI involves giving the language model the ability to think through steps and strategize before executing a task. This is akin to how humans plan and strategize, rather than acting impulsively. In the video, Dr. Ng mentions planning as a key design pattern that can improve the performance of AI Agents. By forcing AI to explain its reasoning step by step, it encourages the agent to plan and think through each step, leading to better outcomes and a more human-like approach to problem-solving.

💡Human Eval Benchmark

The Human Eval Benchmark is a coding challenge or test that is used to evaluate the performance of AI systems in coding tasks. In the video, Dr. Ng uses this benchmark to demonstrate the effectiveness of different AI approaches, including zero-shot prompting and agentic workflows. The benchmark is a way to measure how well AI can solve coding problems that are typically solved by human programmers, providing a practical example of how AI Agents can be assessed and improved.

💡Productivity Boost

Productivity Boost refers to the increase in efficiency and output that can be achieved through the implementation of new technologies or processes. In the context of the video, Dr. Ng suggests that adopting agentic workflows and the use of AI Agents can lead to a significant productivity boost in various applications. By leveraging the power of AI Agents to collaborate, reflect, and use tools, tasks can be completed more effectively and efficiently, leading to improved performance and outcomes.


Dr. Andrew Ng discusses the power of AI agents and their potential to shape the future of artificial intelligence.

AI agents can reason at levels comparable to GPT 4 through the use of models like GPT 3.5.

Agents can transform non-agentic workflows, making them more iterative and effective.

Multiple agents with different roles and tools can collaborate to achieve better outcomes.

Agentic workflows can lead to significantly improved results, as demonstrated by coding benchmark studies.

GPT 3.5, when used within an agentic workflow, can outperform GPT 4 in certain tasks.

Reflection is a powerful tool that allows AI models to self-evaluate and improve their output.

Tool use in AI agents can incorporate pre-existing tools and libraries, expanding their capabilities.

Planning and multi-agent collaboration are emerging technologies with immense potential.

Agents can simulate different personalities or roles, providing diverse insights and solutions.

Agentic workflows can help AI agents recover from failures and improve over time.

The future of AI may involve a shift towards more agentic reasoning and design patterns.

Fast token generation is crucial for agentic workflows, which often involve multiple iterations.

The path to AGI (Artificial General Intelligence) may be advanced through the use of agentic workflows.

Upcoming models like GPT 5 and others are anticipated to further enhance agentic capabilities.

Dr. Andrew Ng's talk emphasizes the importance of embracing agents for a productivity boost and future AI advancements.