I’d never made one before. It lives in Slack now; it knows our social media inside out, and honestly, I mostly just say thank you to it.
Last week someone on the team asked how our Instagram did last month. A month ago, that question would have cost me twenty minutes: open the analytics, export, screenshot the parts that don’t export cleanly, line them up, do the math. This time I typed the question into Slack, tagged a bot, and had the answer before my coffee finished pouring.
I built that bot. I named it Dexter, after the cartoon scientist, because the whole thing started as a tiny experiment and the name made me smile every time it answered. It’s the first agent I’ve ever made, and I want to be clear before I go any further: I am not an engineer. I run marketing. I had never built anything like this in my life until a couple of weeks ago.
Here’s the thing I didn’t realize I’d accepted. Every “quick question” about our social media was slow. Someone would ask which channel was growing, or how last month compared to the one before, or what our best post was, and each of those five-second questions carried a twenty-minute answer buried in three different dashboards. So people mostly stopped asking. In hindsight, that’s the part that actually bothered me. The friction didn’t just cost me time. It quietly made us less curious about our own numbers.
What exists now is Dexter, sitting in our Slack and wired straight into our analytics. I can ask it almost anything. Which channel grew this week. How this month’s engagement stacks up against last month’s. Whether X or Instagram carried our reach. What the single best-performing post was and how many reactions it got. It answers in plain language, live, in the channel where we already talk. No export, no deck, no waiting on me. And it doesn’t stop at the number. If I ask what to do about a dip, it suggests something concrete: lean into the channel that’s climbing, rethink the format that’s flatlining, post the kind of thing that worked last week. Answers plus a next move, in the same breath.
Once the asking worked, the reporting almost built itself. Since the agent already had every number, I taught it to post a full visual report on its own: every channel side by side, each metric compared to the previous period, the wins pulled to the top, and the best posts shown with their images. Each report closes with a short, plain-language read of what the numbers actually mean and what we might try next, so it lands as a recommendation, not just a scoreboard. It drops the weekly one every Monday morning and a longer monthly one at the start of each month, with a short note about the dates it covers. I used to make that by hand. Now I find it waiting for me.
For the curious, here are the guts of it. It’s built on an open-source framework called eve. I stitched it together on my own laptop, in a terminal, and deployed it to Vercel so it runs on its own clock without my machine being on. A month ago, I couldn’t have told you what a single word in that sentence meant.
And here’s what struck me, looking back. Not one of the walls I hit was about intelligence. They were fonts. They were billing. They were a checkbox for a permission I didn’t know existed. The smart part, the part everyone frets about, was the easy part. The work was the plumbing.
That’s the whole thing I’m taking from this. I’d always assumed building an agent was an engineer’s job, something I’d need permission to even attempt. It turns out the barrier was never technical talent. It was just being willing to sit with small, boring problems until they gave way. Intelligence is something you rent now. The patience is the part you bring yourself. And patience I already had. I’d just never pointed it at my own work.
So here’s my question for you. What’s the thing you ask your data over and over, the one that takes five seconds to ask and twenty minutes to answer? I think that’s the one worth handing to an agent first, before the fancy stuff, before anything. Tell me yours. I’m curious whether it looks as automatable from the outside as mine did once I finally sat down and tried.