Companies have extremely unrealistic expectations of artificial intelligence. They think they can bring in a couple of agents, run the tests, and lay off the entire team, saving millions of dollars. I asked my class whether that is realistic right now, and the answer was no; I believe everybody is waking up to how overhyped it is.
A colleague started a company on agents. They got very excited: agents will do all of it, agents, agents. No matter how hard they worked on the design, the agent did ninety-nine things really well and broke one thing big, and that one offset the ninety-nine.
So when you bring an agent workflow into your company, talk to your team first, with confidence. Agents are super capable, but they are not reliable. Eighty percent of the harness is engineering you can copy. Twenty percent is experimentation in this company, and it takes time.
Before the pilot starts, your director asks you to promise one thing: no mistakes in production. Can you promise it?
No, and do not try. Promise what you can keep. The agent will make mistakes, and the harness is built to catch them. Gates stop them before anything reaches production, a sandbox gives it a place to practice, a log records every session, and every mistake becomes a rule so it does not happen twice. Promise that, and give a date for the first numbers: how many mistakes the gates caught this month, and how many got through. A director who hears that on day one is not surprised by the hundredth task. Then put agent harnessing on your résumé, because being able to talk about it and implement it is genuinely a cool skill to have.