Ilya Petrov

My first AI agent

The first alpha of my marketing AI agent is compiling and running (Koog, Kotlin, IntelliJ IDEA, Junie, AIA). So those are my thoughts about its potential. Agents vs “Can I just do everything in ChatGPT”:

  • Feedback loop. The simple thing that agent can run, validate itself, and run again are brining results to the next level, especially with the growing power of big LLMs.
  • Even more thinking. Multi-steps also mean better planning, research, exploration, outlines, testing several drafts, refinements, on a scale.
  • Persistent context. Mix of company’s knowledge and web search are great for both factual precision and exploration of new ideas.
  • Dynamic prompting. Your tweet, blog long read, email and landing could benefit from channel-specific inputs. And so are campaigns for different personas, communication objectives, etc. You can have a hundred prompts ready, but it’s just more convenient to generate it.
  • Result-oriented validation. In coding it’s easier: you compile, run tests – “check”. In a broader marketing sense, the real value is not always measurable on the spot in a straight forward way. Still, in some areas – yes. And also you can rely on historical data.
  • Tooling. Messengers, project tracking, content planning tools – everything is context, and everything is canvas. No more copypasting from emails to Google Doc to blog to reports. But also SEO brains, analytics, and much more.

    The one run of the agent costs me in between $0.005 and $0.05 for the text that may take a human an hour. One hour of copywriter goes from $30 to $300 (some outsource pricing). But it’s not about brining the cost down. Just imagine how much LLMs can accomplish for $30, and what it will mean for this one text! More research, more ideas, more directions, more craft, more compliance checks. And it’s also not about replacing marketeers. Just imagine how much you can accomplish if you are not filling your day as a servant to inefficient processes, moving (virtual) papers from one (virtual) table to another, and not being bothered rewriting the same message for ten different platforms.