top of page

How I Built and Used an AI Agent to Level Up

What Is an AI Agent?

An AI Agent isn’t just ChatGPT with a fancy name. It’s an intelligent system designed to perceive, decide, and act based on goals you assign (Russell & Norvig, 2020). Think of it as a digital teammate—one that listens, remembers, and follows through (without taking coffee breaks).


Unlike typical AI chat tools that respond turn by turn, agents are more advanced. They can:

  • Retain memory across sessions

  • Use tools and APIs to complete real tasks

  • Plan and execute multi-step workflows toward specific goals

In my case, I didn’t just want a chatbot. I needed a project assistant, a thinking partner, and a task manager—all in one. That’s exactly what an AI Agent became.


The first time I used an AI Agent 

In the course named Build AI Application for Education taught by Professor Xavier in Steinhardt in 25 Spring semester. In the course, we learned how to use Langflow+Chainlit to build and publish a personalized AI Agent. For the class final project, we also explored other tools for building AI Agents, like Coze. The underlying logic is consistent, but the user experience of building the agents is better than Langflow, but it is not free.


From trial and error, I learned that the structure of the AI agents does affect the speed of the work; we are supposed to choose a suitable structure for our agents. Like Sequential Architecture for step-to-step in a fixed order, Parallel Structure for multiple tasks or agents run simultaneously, Hierarchical Architecture for high-level agents plan, and low-level agents execute. Additionally, since AI Agents are based on the development of AI, there are still some technical limitations, like image generation and link search. As we know, the related links provided by AI are not available or accurate in most cases.


Pros & Limitations: What Worked and What Didn’t

The Pros

  • Personalization

    • Agents can adapt to your tone, task style, and toolset—basically a mini-you in digital form.

  • Efficiency

    • They automate repetitive actions (e.g., weekly content generation, reminders, research drafts).

  • Scalability

    • Once set up, agents can handle parallel tasks more consistently than I can on a busy day.


The Limitations

  • Setup Time

    • It’s not plug-and-play. You’ll need to test, adjust prompts, and possibly do some light coding.

  • Context Loss

    • Without a good memory structure, agents may "forget" key details mid-task.

  • Hallucinations

    • Like any LLM, agents can generate confident but false outputs. Always double-check critical results.


As Bubeck et al. (2023) observed, LLM-based agents show promise but still face challenges in long-term reasoning and task persistence.


How I Actually Built One (Step by Step)

Here’s how I built mine—no CS degree required:


  • Step 1: Define the Agent’s Purpose

    • Start with a clear, single function. Mine was “instructional design support.”


ree

Figure 1

Screenshot of the DocsBot prompt library homepage showing available features for AI prompt generation. Source: DocsBot.ai, 2025, from https://docsbot.ai/prompts. Screenshot captured by the author under fair use for educational purposes.



  • Step 3: Map the Workflow

    • I used tools like Langflow and ChatGPT API to visualize and connect each part of the agent logic.


ree

Figure 2

Screenshot of a Langflow visual programming interface showing the flowchart for a document-based Retrieval-Augmented Generation (RAG) agent. Source: Screenshot captured by the author from Langflow (https://www.langflow.org), 2025. Used under fair use for educational purposes.


  • Step 4: Test and Refine

    • No agent works perfectly on the first try. I constantly asked: Did this output make sense? Is it doing what I expect?


ree

Figure 3

Screenshot of the Langflow Playground interface showing a dialogue between the user and a custom AI agent (llama-3-8b-instant). The agent responds to a history-related prompt with structured teaching options. Source: Screenshot captured by the author from Langflow (https://www.langflow.org), 2025. Used under fair use for educational and illustrative purposes.


  • Step 5: Publish your Chainlit APP in Hugging Face


ree

Figure 4

Screenshot of a Chainlit application published on Hugging Face, demonstrating an AI Essay Tutor interacting with a user. The chatbot explains its purpose and capabilities, focusing on writing feedback and improvement. Source: Screenshot captured by the author from their deployed Chainlit app on Hugging Face (https://huggingface.co), 2025. Used under fair use for educational purposes.


  • Step 6: Keep It Focused

    • The more specific the task, the better the result. Overloading an agent makes it weaker, not smarter.


Pro tip: Complexity kills. One focused agent > one confused generalist.


What’s Next: Lessons, Pitfalls, and Possibilities - What I’ve Learned

  • Start small: Build one workflow first, then scale.

  • Be clear: Know exactly what you want your agent to do.

  • Prompting is design: A good prompt is more valuable than a fancy interface.


Pitfalls to Avoid

  • Tool overload: Too many platforms = confusion. Pick one and master it.

  • Excessive context: Agents lose focus when fed too much.

  • Blind trust: Always review critical outputs—hallucinations happen.


What’s Next

I’m now working on agents that connect to live tools—like Notion, Google Calendar, and email. I also want to build a self-reflection agent to help me journal, spot patterns, and support personal growth over time.


This journey taught me one powerful thing:

AI Agents don’t replace you—they amplify you.


References 

Russell, S., & Norvig, P. (2020). Artificial intelligence: A modern approach (4th ed.). Pearson.


Bubeck, S., Chandrasekaran, V., Eldan, R., Gehrke, J., Horvitz, E., Kamar, E., ... & Zhang, Y. (2023). Sparks of artificial general intelligence: Early experiments with GPT-4. arXiv preprint arXiv:2303.12712. https://doi.org/10.48550/arXiv.2303.12712

DocsBot.ai. (2025). Prompt Library. https://docsbot.ai/prompts


Langflow. (n.d.). Langflow: Build LLM applications visually. https://www.langflow.org


Hugging Face. (n.d.). Hugging Face – The AI community building the future. https://huggingface.co


Chainlit. (n.d.). Build Python LLM apps in minutes. https://www.chainlit.io/


 
 
 
bottom of page