Why AI Agents are a Hype (Still)

What’s an AI agent? … A BUZZWORD.

Well, not really. The stuff that rebranded AI consultants call “agents” are part of a hype.

True AI agents are still a far cry away – autonomous systems that reliably plan a set of steps in order to achieve a goal. And still be reliable when the goal changes every time. Something like “draft a proposal for client X considering market Y in the last year and our customer conversations”.

There are approaches to this, but they are still in experimental stages at the time of this writing. They still make tons of errors.

When you hard-code (i.e. deterministically specify) the steps of a process, there’s no planning involved. Planning involves creativity.

What these AI cowboys sell you are LLMs with tool access. That means that the LLM (e.g. ChatGPT) formats the output in a machine-friendly way. Then a program executes it behind the scenes.

For example if you tell an LLM with tool access “book an event in my calendar for 12/1 at 10 am to get my haircut”, what it actually does is to translate that sentence into something like:

POST /api/v1/calendar/events

{
  "title": "Barbershop appointment",
  "description": "Getting your haircut",
  "start_time": "2025-01-12T10:00:00Z",
  "end_time": "2025-01-12T11:00:00Z",
}

and execute that on a backend system. The backend system then makes the booking in your calendar app.

That is syntactic transformation. It has no new meaning, just a different way to write it. 

No planning. No creativity. No problem solving.

Now this doesn’t mean that these tools aren’t useful. They’re a GAME-CHANGER, because never before were these transformations possible at this complexity!

I’m only saying that rebranded AI cowboys don’t know what they’re talking about:

But that doesn’t matter. Their hype raises awareness and adoption in the broader industry, which is the hardest part.

When applied smartly, LLM tool usage in combination with smart software to coordinate it can create some truly magical automations! 

By the way, probably around 90% of business problems can be solved by optimization algorithms developed during the 1950s and 60s. And the power of real agentic workflows comes from the interplay between exact algorithms and LLMs. As LLM capabilities are getting better, collaboration between different agents seems promising – which is also in active research right now, and not business-ready.

But AI cowboys don’t know that. How can they?

Their job is to sell you whatever the “next thing” is – usually wrapped in a buzzword – without going deep into the subject matter. That’s easy.

But the hard part is diving deep into a business problem and choosing the right tools to solve it, based on their technical capabilities and suitability.

If AI development were to stop now, we’d still have ten years worth of innovation to properly make use of in a business context. The major breakthroughs have already been made as far back as March 2023 (at least!), when OpenAI released the “tool usage” capabilities for their chat models.

But it took almost two years for these to reach the broader business public. So in a way the AI cowboys think they’re riding the newest hype, but in fact it’s 2-year-old tech. That’s just to point out how far away from actual tech expertise they are.

If you want to learn how to use Generative AI, tool usage, and maybe even wander into the AI agent paradigm, we’ve you covered in the Make Work Obsolete community.

You’ll be guided towards building your first AI automation by real tech experts, without breaking the bank. Join weekly Q&A calls, a forum with other community members, and a full AI course curriculum. Try it for FREE for 7 days.

We’re giving away 50% off subscriptions for the readers of this newsletter until end of January with code START2025. So be quick to redeem your offer!

Cheers,
Robert