min. readSo, you’ve got agents running. Maybe three, maybe seven, maybe someone on your team just dropped a Notion doc titled “Agent Strategy Q3” and now there’s a meeting about it. The question nobody is asking out loud yet is who’s actually watching these things.
Here’s what agents are, stripped of the hype: each one does one job. It handles a specific input, runs a specific process, and spits out a specific output. That’s how they work, by design. And every single one of them needs a human behind it who set it up, keeps it calibrated, and steps in when it starts doing something technically correct but completely wrong.
Think about it this way: would you give an agent full access to your bank account with nobody in between? No one reviewing what it approved, no override button, just the agent doing its thing with your money. That question makes most people uncomfortable, and it should. But somehow that same level of trust gets extended to agents running inside workflows that are just as sensitive, with way less oversight than you’d give a new hire in their first week.
Demos are easy. Production is where things get honest.
Every agent has had a great demo. Clean data, narrow scope, the right people in the room nodding along. Then it hits a real environment with fragmented systems, three different naming conventions, and an API that someone deprecated six months ago but forgot to mention to anyone.
In 2025, 46% of AI pilots got scrapped before reaching production, and nearly two-thirds of companies were still stuck in proof-of-concept. Agility at Scale Technology wasn’t usually an issue. Getting agents out of the sandbox means someone has to do the unglamorous work of connecting them to real infrastructure, building the logic that ties them together, and writing the guardrails that keep them from confidently breaking something on a quiet Friday night. That work requires engineers who understand the full picture, not just the model.
The more agents you run, the more humans you need.
This is the part that gets left out of the pitch decks. Agents don’t scale themselves. They need to be customized for your context, monitored for drift, updated when the world changes, and occasionally overruled by someone with actual judgment. Scaling agentic systems for real requires people who can hold machine learning, data engineering, systems integration, and AI governance in their heads at the same time. AkrayaThat’s a specific kind of team, and most companies don’t have one sitting around with the capacity to spare.
So the problem most engineering leaders run into isn’t that agents don’t work. It’s that building the human layer around them takes a kind of talent that’s genuinely hard to hire for, slow to onboard, and expensive to keep if your existing team is already stretched. You need people who can move fast, communicate without handholding, and understand your stack well enough to make judgment calls on their own.
Bain found that companies that have scaled AI across their workflows are already posting EBITDA gains of 10% to 25%. Bain & Company The difference between those companies and everyone still running the same pilot from eight months ago usually comes down to whether they had the right people around the technology, not the technology itself.
Running agents without the right team is like opening a restaurant where the kitchen is fully automated, but nobody trained the machines, nobody checks the orders, and the chef is on a different floor answering emails. The food might come out fine. Or it might not. Either way, someone is going to have a bad night.
People carry the responsibility and keeping them in the loop is what will move this to the next level. When agents operate on their own with binary, black, and white thinking, they can make decisions that affect things in ways we cannot fully predict.
About the author:
Mauricio Goita is Lead Generation Manager at Abstra. He built his career in tech after originally graduating as a lawyer, a path he chose not to pursue professionally. Since entering the industry in 2018, he has grown from junior roles into commercial leadership, contributing to major revenue growth and building teams from the ground up. Today, he is focused on automation, demand generation, and the use of AI in internal processes.