ReAct vs Plan-and-Execute: The Architecture Behind Modern AI Agents
In this video, we explain the patterns powering reasoning models in 2026
Good morning!
AI agents are evolving really fast, and the real shift is not just bigger models; it is a better reasoning structure that we can build around as AI engineers.
Traditional LLMs treat complex tasks like single prompts. They create tokens from tokens. They live in the text world. They generate something that sounds coherent, but without real planning, verification, or access to tools that allow them to take action or adapt in real time.
On the other hand, modern agent systems change that by separating thinking from acting, allowing models to plan steps, use tools, observe results, and adjust along the way.
Two key patterns drive this shift: ReAct and Plan-and-Execute.
ReAct loops through thought, action, observation, which makes it powerful for messy and uncertain tasks where new information keeps changing the situation.
Plan-and-Execute commits to a structured roadmap upfront, which makes it more efficient for predictable workflows. The most advanced Deep Research systems combine both by using high level planning with local adaptive reasoning loops.
The big takeaway is simple. Building strong AI agents is not about adding more intelligence. It is about choosing the right level of structure (autonomy vs. workflow-level control) for the uncertainty you are dealing with, and designing systems that know when to adapt versus when to commit.
In this week’s video, I break down how these architectures actually work, and when to use each!
By the way, this video is one of the few free lessons we have in our brand new Agentic AI Engineering course we just released! If you are interested in learning to build agents the right way, not just use CrewAI and such other systems, but build sustainable ones for your products or company, check out our new course here! Since you are a follower of the content, use code “louis_supporter” to get 15% off the full price, right now at $499, for the next 96 seats!
Watch this week’s video on How AI Agents Actually Work: ReAct vs Plan-and-Execute:
Your takeaway about choosing the right level of structure for the uncertainty you're dealing with is spot on. Jeff Dean frames it slightly differently. He calls planning "the bridge between a model that can answer a question and a model that can actually accomplish a task." Same insight, different angle. I covered his full take on it here: https://reading.sh/jeff-dean-on-what-actually-makes-ai-agents-work-dced5bb50206?sk=d8b9e7faac0da6011382834459ca4808
The framing of 'autonomy vs. workflow-level control' is where this gets practically useful. Most tutorials skip straight to implementation and miss the architectural question entirely. The part I'd push back on slightly: in my experience, even 'predictable' workflows turn messy the moment real data or edge cases hit them.
ReAct's adaptive loop ends up saving you even when you planned for structure. The course looks solid - this is one of the cleaner explanations of why the architecture choice matters before you reach for a framework.