When I get a new toy tool from the hardware store, I generally love to find as many use cases for it as possible and get the most out of my new purchase. Businesses felt exactly the same way when AI became accessible and packaged up for them as “agents”. And so the great agentification began: suddenly every business wanted agents.
So what’s an agent? It’s simply asking an LLM to behave like a human when given a certain task. So the agent will then process information, make decisions, use tools etc., very much like a human employee would, but, in theory, much faster and cheaper (and arguably better). Agentification is creating agents for a whole variety of tasks. Agentic is the way of thinking that a problem or task should be handed over to an agent.
Agents are great in the same way humans are: flexibility, and (for most humans) the ability to adapt to not-so-perfect contexts and situations. In essence the very opposite of what software is great at. Code defines a precise context and set of parameters and a specific logical flow to follow in order to achieve the desired outcome.
OK, so here comes the catch. Agents sound pretty awesome, right? So why not have them everywhere?
Well, for two reasons:
- It’s like killing a mosquito with a shotgun. It will get the job done but… really?
- The agent’s flexibility is nice, but you know what else is nice? Deterministic outcomes! That is, knowing that when given the same task again the outcome will be the same. Agents, in their human way, are not very good at consistency. Just like humans they can get confused when given ambiguous parameters and also have bad days where what used to be simple is now confusing. We can sometimes circumvent this by running the same task multiple times, comparing outcomes and taking the majority answer, but that gets clunky real quick and increases cost while slowing the process. Not ideal.
So it turns out software is still pretty good at handling most everyday work tasks, and it does them consistently, quickly and for next to nothing. What’s changed is that the software itself is now almost free to write, so let an LLM build the code and the infra. Then sprinkle agents on top, where the work is messy: reading the brief, turning it into parameters, catching the exceptions nobody thought to specify. Agents to understand the job. Software to do it.