Do You Want to Build Useful AI Agents?
Slow down cowperson, take a breath, and relax
Now that we are calmer let’s approach this project rationally. AI agents are all the rage right now. That might change in a day or two, but let’s pretend they’ll be around for a while. You’re excited to build an AI agent to pilot the space station autonomously. Cool! I love it! Let’s do it! Wait, and let me take a breath now. Okay, let’s follow a simple checklist I’ve put together for planning out your agent. Why should you listen to me? You shouldn’t, but I’ve been building software since I was 8 (many decades ago). Some of it has even been useful. I do know some stuff. Here we go.
Tip #1 Define Your Agent’s Purpose
Define your agent’s purpose and, more importantly, the need it fills. I see a lot of people rolling their own everything with AI, and typically, I support that, but we do not need 8,000 customer service bots that essentially do the same things in the same ways. If it doesn’t serve some novel purpose, please keep going until it does. Obviously, you can do what you wish, but this is my checklist, so I want to see creativity.
Tip #2 Data Quality
Feed your agent good data. In the years I worked with the Splunk platform, that was the number one problem we had to overcome. No one ever had good, clean data we could use. We had to pull teeth to get anything of quality into the engine and get good results out of it. I often refer to an old computer science adage, GIGO, Garbage In — Garbage Out. There’s almost nothing more accurate in the world. You could easily apply to education or the law or anything. If you educate a child poorly, they will spew a lot of garbage. It’s the same with AI.
Tip #3 Don’t Make A Dumb Agent
Can you imagine spending a ton of time training an agent only to have it not be helpful a week later? That can happen in business quickly. You must ensure your agent’s knowledge is constantly (well, nearly continually) refreshed. Imagine you’re running a 35% deal on your website, but your chatbot has no idea about the promotion. A user might grow very frustrated. You’re making them talk to a robot, and that robot doesn’t have the relevant information. That’s a customer service disaster waiting to happen. This brings us to tip number four.
Tip #4 Focus On The User Experience
You’ve spent however long building out this fantastic agent that will revolutionize your customer service. There’s just one problem: your agent is a jerk. You probably want your agent to have a pleasant personality and also be helpful. You want to help your agent help your customers, so try to prepare it for your business's everyday situations. This is an often overlooked part of agent building. I especially like to approach training chatbots as I would a live customer service person. They mustn’t break down when asked a slightly more advanced question than “Hello.” I encourage business owners always to try to see how two other people use and see things. The first is your employees, and the second is the customer.
Tip #5 Adversarial Testing
This is my last tip for this list, but it might be the most important. Test your agent. Try to make it do things you wouldn’t want it to do. This is key in any software endeavor. It will boost both the quality of your agent’s response and the robustness of your agent software. This step gets skipped often, but I encourage everyone to do it. Think of it as penetration testing for your new AI best friend. It really is crucial.
Conclusion
These are just my top five tips for building a good AI agent. They’re honestly essential tips for any software development and are often overlooked. They can apply to anything you’re building.
