# Choose an AI workflow or agent by the decisions the job requires Use the three-level explainer to define model work, tool actions and the controls around a complete task. [HTML](https://lukejbyrne.com/library/guides/ub52ra38wxc/) Content language: en Author: Luke Byrne Video published: 2025-10-16 Guide published: 2026-11-02 [Watch the video](https://www.youtube.com/watch?v=Ub52RA38wxc) Video language: English ## Separate the model, workflow and agent roles Luke’s October 2025 explainer distinguishes LLMs, workflows and agents as three levels of automation. It introduces ReAct, the TAO loop and frameworks including LangChain, CrewAI and Autogen, alongside MCP. These terms describe different parts of a system; naming a framework does not show that a particular task has been solved. 1. Choose one job and list its inputs, desired output and available tools. Mark which steps simply transform supplied information and which need access to an external system. 2. Identify where a fixed sequence is sufficient and where the next action genuinely depends on a result the system must inspect. Use that decision point to explain why the job needs an agent rather than adding autonomy by default. ## Make actions and observations visible The agent section begins at 3:47 and the workflow-versus-agent discussion at 5:00. The useful distinction is what the system can do with a response: suggest an action, execute a permitted tool call or observe its result and choose another step. A fluent chat answer is not evidence that the external task happened. 1. Write a small acceptance example with the expected action and observable result. Include a case where the tool fails or information is missing so the system has a defined response beyond repeating the same request. 2. Keep the input, selected action and returned evidence together in the run record. Inspect the actual saved file, record or destination rather than accepting a completion statement alone. ## Choose implementation tools after defining the limits The framework overview starts at 5:33 and pitfalls at 6:08. The description highlights infinite loops and unexpected costs. The retained learning resource is https://datacamp.pxf.io/19vVr9. Use the framework discussion to organise implementation choices, not as a guarantee that one library prevents every failure. 1. Define a practical work or usage limit, the allowed tools and the actions requiring human review. Give the system a clear stopping condition when the task is finished, blocked or no longer producing useful progress. 2. Test a small run and inspect its actions, result and usage before expanding access. Compare it with a simpler fixed workflow so you can see whether the added decision-making actually improves the job. ## The takeaway Use autonomy where a task needs decisions based on observations. Keep tool access, stopping conditions and evidence of completion explicit whichever framework implements the system. ## Workflow resources - [datacamp.pxf.io](https://datacamp.pxf.io/19vVr9) - [ailuke.short.gy](https://ailuke.short.gy/Ub52RA38wxc1) - [ailuke.short.gy](https://ailuke.short.gy/Ub52RA38wxc2) - [datacamp.pxf.io](https://datacamp.pxf.io/yqr3jb) - [newsletter.lukejbyrne.com](https://newsletter.lukejbyrne.com) - [linkedin.com](https://linkedin.com/in/lukejbyrne) - [x.com](https://x.com/thelukejbyrne) ## Chapters - [0:00 – What Are AI Agents?](https://www.youtube.com/watch?v=Ub52RA38wxc&t=0s) - [0:22 – LLMs (Level 1)](https://www.youtube.com/watch?v=Ub52RA38wxc&t=22s) - [1:08 – AI Workflows (Level 2)](https://www.youtube.com/watch?v=Ub52RA38wxc&t=68s) - [2:26 – Deliver real value with AI agents](https://www.youtube.com/watch?v=Ub52RA38wxc&t=146s) - [3:47 – AI Agents (Level 3)](https://www.youtube.com/watch?v=Ub52RA38wxc&t=227s) - [5:00 – When to Use Agents vs. Workflows](https://www.youtube.com/watch?v=Ub52RA38wxc&t=300s) - [5:33 – Frameworks: LangChain, CrewAI, Autogen](https://www.youtube.com/watch?v=Ub52RA38wxc&t=333s) - [6:08 – Common Pitfalls in Agent Design](https://www.youtube.com/watch?v=Ub52RA38wxc&t=368s) ## Sources and limitations Prepared 2 November 2026 from the published video description and chapters. The actions are editorial follow-along guidance. The full video was not independently watched or transcribed and no new product test was run. Setup, availability and offers may have changed since the recording. The levels are the video’s teaching framework, not a universal classification of every product called ChatGPT or an agent; no current framework installation instructions are supplied. [Original video](https://www.youtube.com/watch?v=Ub52RA38wxc)