I built a simple AI agent that calls the PostNL API to get package tracking info.
Here is the LangFlow diagram:
I live in the Netherlands and use PostNL a lot.
When there’s a problem with a package, the process is too complicated. You have to call customer support, wait a long time, read help pages, and check tracking links. It’s too much work for simple questions.
So I asked myself: what if one AI agent could answer all questions about sending or receiving a package? This question started the project.
The Goal
The agent should answer questions about:
- Sending a package
- Receiving a package
- Delivery delays or plan changes
To fully cover the PostNL website, the agent would need a RAG system, meaning it can read all help pages.
But for real-time tracking, it must call APIs, because website content alone isn’t enough.
I decided to start simple. The first version has one main job: call the PostNL API and get tracking info.
Building the Solution
I built the agent using LangFlow. It’s more than a bot, it remembers previous queries, calculates delivery times or costs, calls the PostNL API for live tracking, and logs every step for improvement.
The agent uses two main tools:
- Calculator – handles time, money, and other calculations.
- Custom API tool – a LangFlow node that fetches real tracking data from PostNL instantly.
With these tools and a memory, the agent can answer questions using saved data, avoiding repeated API calls and making responses faster.
Testing the Agent
I tested the agent in a playground to ask real questions and see the answers.
Testing early helped me catch problems quickly.
Debugging with LangSmith
I set up LangSmith tracing to see how the agent works step by step.
This helped me:
- Track each step of the agent
- Find wrong answers
- Improve prompts and tool usage
- Understand failures
Without tracing, debugging would be much harder.
Key Takeaways
This small project shows how AI agents can reduce user effort.
Next steps are adding RAG and covering more use cases.
If you’re an AI product manager, start with small agents. You’ll learn faster than just reading theory.