AI Summary
5 min readShawn Kim, head of Morgan Stanley's Europe and Asia Technology Team, outlines a shift in AI from generative models that assist with thinking to agentic AI that takes action. In this solo episode of Thoughts on the Market, he explains how this transition alters computing demands and creates investment opportunities in overlooked hardware.
From Passive Responses to Active Workflows
Current generative AI, or Gen AI, operates passively: it responds to a user's prompt with text, like a summary or draft, but requires constant human oversight—asking, refining, copying, and checking. Agentic AI moves beyond this to active systems that act independently. These agents remember past interactions, grasp user preferences, integrate across digital tools, plan multi-step workflows, and adapt to changes.
The key distinction lies in usage. Gen AI serves as a copilot for single tasks, while agentic AI functions as an autopilot for complex, ongoing processes. This evolution demands different computing resources. Gen AI relies heavily on GPUs—graphics processing units—for parallel processing in large language models (LLMs). Agentic AI, however, elevates CPUs—central processing units—which coordinate tasks, manage orchestration, and link to broader infrastructure.
The Three Stacks of Agentic AI
Agentic AI rests on three layers: the brain, orchestration, and knowledge.
Continue reading the full summary in the app — free to try.
Read Full Summary →Free • No credit card required
Never miss an episode of Thoughts on the Market
Get every new episode summarized in your inbox — free, ~5 minutes to read.
No spam. Unsubscribe anytime.
What you'll learn
- 1 (00:00) **Episode Intro** - Shawn Kim welcomes listeners and teases AI shift from thinking to action
- 2 (00:17) **Chatbot Example** - Describes current Gen AI as user-driven prompting and checking
- 3 (00:34) **Vision of Agentic AI** - Outlines active AI that remembers, plans workflows, and adapts across tools
- 4 (00:52) **Gen AI vs Agentic AI** - Passive prompt-response vs active multi-step autopilot
- 5 (01:12) **Hardware Shift** - GPUs for Gen AI thinking; CPUs gain for agentic coordination
- 6 (01:43) **Agentic AI Stacks** - Brain (LLM), orchestration (CPU), knowledge (memory)
- 7 (01:55) **Memory's Key Role** - Builds context flywheel for personalization and retention
+ Full timestamped outline available in the app
Show Notes
Our Head of Europe and Asia Technology Research Shawn Kim discusses AI’s move from passive chatbots to active agents—and how this influences tech supply chains.
Read more insights from Morgan Stanley.
----- Transcript -----
Welcome to Thoughts on the Market. I’m Shawn Kim, Head of Morgan Stanley’s Europe and Asia Technology Team.
Today: A foundational shift in the development of AI and its broad market implications.
It’s Tuesday, May 5th, at 3pm in London.
Think about the last time you asked a chatbot to write a summary or a draft. Or maybe answer a query. It was probably useful. But you were also still driving the interaction: asking, refining, copying, checking, and moving the work forward.
Now imagine a system that does not just respond, but acts. It remembers what you asked last week, understands your preferences, works across digital tools, plans a workflow, and adapts as circumstances change.
That is the shift from GenAI to agentic AI: from AI that helps with thinking to AI that helps with doing. GenAI is mostly passive. It takes a prompt and produces an answer. Agentic AI is active – less a copilot for one task but an autopilot for multi-step workflows.
The distinction is key because computing requirements are changing. In GenAI, large language models and GPUs handle much of the thinking. GPUs, or graphics processing units, process many calculations in parallel, making them central to modern AI models. In agentic AI, CPU becomes more important. CPUs, or central processing units, coordinate tasks and connect systems to the broader digital infrastructure.
Agentic AI also depends on three stacks: the brain, or the large language model; orchestration, where the CPU manages the doing; and knowledge, which is memory.
Memory may be the most important layer. An agent that knows your preferences, documents, tone, and task history becomes more useful over time. That creates a context flywheel. The more context it collects, the more personalized it becomes, and the harder it is to leave.
Typically, in computing, we think of memory as storage, mainly. We need to rethink this. Memory is also continuity. When an AI system can use past experiences, memory becomes a long-term state, shared knowledge, and behavioral grounding.
And that matters because LLMs have fixed context windows. Once a conversation exceeds that window, older content falls off. For simple questions, that may be fine. But for a coding agent working across a large codebase over days or weeks, it is a major limitation. Serious work requires persistent memory, short-term orientation, and active retrieval – remembering prior decisions, understanding changed files, and finding relevant codes without the user pointing to every dependency.
For investors,
More from this podcast
Thoughts on the Market →