I Built my First AI Agent: Here are 7 Lessons I Didn’t Expect6 min read
AI agent demos always look slick, and I often come away thinking about the endless possibilities they could lead to. However, it’s very easy to nod along and think, "yes, that’s powerful", without ever quite getting your hands dirty.
So I decided to stop observing and try it, to understand what it really takes to build an AI Agent. These are the nine lessons I learnt.
1. It's More Accessible Than you Think, but not Effortless
The first thing that surprised me was how approachable it felt. I am in no delusion that I am a tech wizard, so I was very smug to find out that there is no heavy coding involved. There was no writing complex scripts of jumbled-up symbols and letters. It could all be configured in plain English by simply telling the agent what to do.
Although, sadly AI is still not able to read my mind, so it didn’t just work. It was not “click a button and you’ve got an intelligent system.” I still had to think carefully about:
- How it should behave
- What it should do
- How it decides things.
That’s where the real work is.
2. I Wasn't Building Just one Thing, I was Building a Team
The agents have to be broken down into parent agents and child agents. The responsibilities are divided by agent, much like the team within a restaurant.
The parent agent is the one having the conversation, like the Front of House and the waitstaff. They are the ones who interact with customers and help action the requests. The parent doesn’t do everything itself; it decides who should handle the request and sends it in the right direction. The child agents sit underneath, each with a very specific job. These are like the rest of the team in the restaurant: the chefs, the dishwashers, the accountant etc. They each have their own specialist roles that all contribute towards the restaurant as a whole, but they never serve the customer the meal themselves.
Building an agent felt more like organising a team than building “AI”.
3. Descriptions and Instructions are Your Guardrails
The parent agent decides where to send requests based on how you describe your child agents. If those descriptions are vague, it gets confused, and when it gets confused, it still tries to answer. This means you might be sending your meal request off to be cooked by the accountant.
It made me realise how important it was to be as descriptive as possible, both when describing the agents themselves and when prompting.
You’re not just telling the agent what to do, you’re shaping how it behaves. Unless you tell it you want your pizzas round and flat, it may serve them cubed. This is why it’s really important to define what rules it should follow and how you want the output to be formatted, otherwise it will fill in the gaps itself.
4. Your Data is Everything
An agent is only as good as the information you give it. If you connect it to a knowledge source, it will use that. But it won’t question whether it should use it... This was probably the most important part of building the agent, because if you point it at the wrong thing, it will confidently surface the wrong information, and can even cause security problems if it offers up confidential information.
5. What Didn't Work and What Caught me Out
This was my favourite part, the bits they did not show in the demos.
Not all models behave the same.
I assumed once something worked, it would just… work. But different models handled things very differently, a bit like giving chefs different cuisines. They may be in the same kitchen using the same tools, but their dishes will likely come out very differently.
One of the AI models I initially used struggled with reading a spreadsheet entirely. Another handled it without an issue. Building the agent wasn’t just about the design, choosing the right model and testing it properly also made a huge differently.
6. AI Isn't the Answer on it's Own
One of the biggest takeaways for me was that AI doesn’t replace everything else. It complements it. There’s a temptation to think agents should automate everything, but that gets expensive quickly (and will use up a lot of tokens), when it’s not always necessary.
Some things are better suited to structured automation using Power Platform, whilst some things benefit from AI flexibility. The real value is when the two work together.
7. Where it Starts to get Exciting
In the future I would love to build slightly more complex agents which can do things like:
- Route queries intelligently
- Recommend the right people to handle a task
- Pull in information from multiple sources
- Support workflows instead of just responding to prompts
What I’d do differently next time
Looking back, there are a few things I’d change straight away:
- I’d spend more time defining the use case upfront
- I’d be much clearer and more detailed in how I describe each agent
- I’d test models earlier instead of assuming they’re interchangeable
- I’d be more selective about what data I connect
- I’d break things into smaller, more focused agents rather than trying to make one do everything
Basically less trying to make it clever, and more trying to make it useful.
Final Thoughts
The first agent you build probably won’t be the one you keep.
But it will change how you think.
Once you’ve built one, even a rough version, you stop seeing AI as this abstract thing that “does stuff”, and start understanding how it actually behaves.
That’s where it gets interesting.
Next Steps
Many organisations are only scratching the surface of what's possible with AI. Whether you're looking to improve adoption, help users get more value from the tools they already have, or explore opportunities to build custom agents, having the right guidance can make a significant difference.
Our Enhanced Support service gives you direct access to experienced consultants who can help your team develop practical Copilot skills, identify high-value use cases, and build confidence in using AI across the business. We can support everything from day-to-day Copilot usage and best practices through to designing and implementing custom agents tailored to your organisation's needs.
If you'd like to discuss how we can help your organisation maximise the value of Copilot, get in touch with sthe Cloud9 team: sales@cloud9insight.com.