There is a particular kind of project that starts with a platform and goes looking for something to point it at. Agents attract this more than most technology, because the demo is so good that it feels like the hard part is already done.
My own version of this was a LinkedIn content agent. The plan was that I would talk to it, and posts would appear. No drafting, no review, no queue. I got far enough into building it to find out that the parts I had waved at were the whole job: knowing which of my ideas were worth writing up, holding my actual voice rather than a generic one, and knowing when a draft was finished.
None of those failed because the agent was not capable enough. They failed because I had never written down what “good” meant, so there was nothing to automate against.
So I stopped and did the boring version first. A Claude project, the steps listed out, run by hand until the output was consistently something I would publish. That took a few passes. What came out of it was a specification, which is the thing I had been missing at the start.
Automation after that was mechanical. The steps existed, so they could be handed over one at a time, and I could tell immediately when one of them regressed because I knew what the output was supposed to look like.
The order matters more than the tooling. A problem you understand can be automated on almost any runtime. A problem you have not defined will defeat the best agent framework you can find, and it will take you a few weeks to notice.
This clip is from my conversation on the 2Stallions Marketing Expert Series.
