The Creator Agent: Build Tiny Tools Instead of Buying More SaaS

The next era of business software is small, custom tools that do one job for one team. You point an agent at a real pain point and it builds you a tool that solves exactly that. No more cramming your whole company into Salesforce, Oracle, or HubSpot. We call the thing doing the building the creator agent. You use it to make tools, and a lot of those tools never touch AI once they run.

In this episode, I talked with Mike Carlo, who runs PowerBI Tips and co-hosts a new podcast called Agentic Thinking. We have been calling this shift hyper-personalized tooling for a while now. Mike put a name on the engine behind it. The creator agent is an agent you talk to like an engineer, and it writes the small programs that run your work. Some run on normal hardware, some are little scripts, and almost none of them need tokens once they are built.

Here is what this looks like with real clients. The scope is changing. People come to us with problems now, not requests for one big workflow automation in n8n or Make.com. We are losing time reading invoices that get emailed to us. Our accounts payable and receivable have SKU mismatches we fix by hand. We keep losing data. So we go in, find the painful spots, and build a tool that does that one thing. One client found out they had $16,000 in inventory sitting on a shelf, 90 days old, that nobody was moving. A small tool surfaced it. They could have cleared it out months sooner.

The speed is the part that still gets me. Mike needed to strip the backgrounds off 80 images. The old path is Adobe or Lightroom and their algorithms. He did not want that. So he told his agent, here is a folder of raw files, build me something that removes the backgrounds, keep the same folder structure, output PNGs. He did not even pick the language. Thirty minutes later he had a tool he can throw any number of images at, forever, with no subscription and no shady website. I had a similar one. A client had thousands of Stripe transactions to match against student records, and some were missing IDs or emails. I dropped a 55,000-row spreadsheet into Claude Code, explained the mismatch, and it wrote a little Python script. It tested on a few rows first, found 15 matches out of the 16 it tried, then ran the whole file in about 3 seconds.

The reason this matters for a regular business is that the hard part has moved. The agent already knows more Python and JavaScript and Rust than I will ever have time to learn. My senior engineer has not written a line of code by hand since October. Where agents are weak is system design. You can tell an agent "I have problems" and it has no idea what to solve, because it does not know your inputs, your outputs, or the way your workflow runs. So the skill now is stitching small tools together into one good process. That is your secret sauce. Mike said it well. Many people can do what you do, and if your process is a little better, that edge stays yours for a long time.

We also talked about why you should look at boring software. Mike pointed to a clip from Jared Friedman at Y Combinator about the SaaS moat going away. Cloning a popular CRM to resell it is a crowded race nobody wins. The money is in the niche tools no one has touched. Picture a CRM built only for commercial kitchen hood exhaust cleaning companies. Or industrial control systems with one or two providers that have never changed because there was no competition. Most off-the-shelf software does 80 percent of what you want and skips the 20 percent that makes it useful. The tighter and more specific you go, the more valuable the tool becomes.

One thing I want owners to hear. Do not be scared to keep simplifying. Since building is so fast now, you can ask the agent to make a tool simpler, then ask again tomorrow. Elon has a line I like about deleting so much from a design that you have to add 10 percent back, because if you never overdelete, you are not deleting enough. Karpathy does the same thing, boiling a model down to a couple hundred lines you can read by eye. Use source control, push to GitHub, and keep asking the agent to trim. Fewer parts means fewer points of failure.

At the end of the day, you do not need a full redo of your systems. You need to find the specific spots where your team loses time, makes errors, or drops data, then build a small tool around each one. Some will use AI, some will not. Either way they make the day-to-day easier and the business more efficient. Your competitors are reaching for the same agents. So you either use AI in production or you use it to build the tools that run your shop. Both beat paying for one more subscription that almost fits.