Agents Are for Exploration
It is easy to look at AI and see a faster worker. You feed it a job, it gets the job done, and perhaps "cheaper". Then comes the disappointment. It always does. The agent does not run the job exactly the way it was written down, so the technology must not be ready. My argument is that this is the wrong perspective, and it will not get less wrong as the models improve.
That picture is the factory line. A worker performs one step the same way every time, and you grade them on how exactly they hit the spec. Hand AI that and yes, it wobbles hard. It is not deterministic enough to follow a fixed recipe to the letter. But the work we hand it is not factory work. The jobs that matter need judgement and freedom, and there are ten valid ways to reach the same end. We drop AI into that messy middle and then scold it for not behaving like a machine on a line.
What I am proposing is that AI is good at exploration. A task that can be done ten different ways, with some margin for error, is a great fit. A task that has to follow one exact recipe is not, and if you already have the recipe you did not need an agent. You needed a script. Write the script!
Every success I have seen comes from pointing an agent at a fuzzy problem instead of boxing it into a rigid shape. Our internal SDR agent does not blast outreach and spam people in a loop with a fixed outreach pattern. It is figuring out our ICP, running experiments, reading the data, taking the signals we hand it, searching for an answer we cannot write down ourselves. Our support agent does not exist to close tickets. It chases us, nags us, even calls us, so we get better at support. It is hunting for a better process. The point was never to automate away a conversation nobody wanted automated.
We just put a few of these in the open where you can watch them work. The paper trader has no strategy baked in. It is looking for one. If we had a strategy we would have coded the algorithm. We do not, so we sent an agent to find it. The coder decides for itself what is worth building. The arcade ships a brand new game every single day, and I genuinely do not know what tomorrow's will be. I could hand it a recipe for a game. That would be boring. Instead it goes looking, and I check in to be surprised.
That is the whole shift. Automation uses AI to do what you already do, only faster, which crams a new capability into an old shape. Exploration uses it to find solutions you would never have reached on your own. Only one of those is worth getting out of bed for.
If the work has a recipe, write the script. If it does not, that is exactly where an agent earns its keep.