Your first agent should embarrass you

The most common way an AI project dies is not failure. It is ambition. Someone maps the whole department, draws the agent that will read every inbox, reconcile every record, and answer every question, and then spends four months building a version of it that is never quite ready to let near a real customer. The plan is gorgeous. The org chart of agents is gorgeous. Nothing runs.

Meanwhile the agent that would have actually changed something this week was small enough to be embarrassing. It does one thing. It reads the support queue and tags each ticket by product area. That is it. No reasoning chain to brag about, no swarm, no diagram. It would lose a demo to the grand plan in front of any audience. And it is the one that should ship.

Small is not a phase, it is the strategy

The instinct is to treat the tiny agent as scaffolding, a throwaway step on the road to the real thing. That instinct is the bug. The tiny agent is not a worse version of the grand plan. It is a different bet, and a better one, because it pays you back in the only currency that matters: contact with reality.

A one-tool agent in production teaches you things the plan cannot. It tells you the tags are wrong half the time because two product areas overlap in a way nobody documented. It tells you the inbox has a category you forgot existed. It tells you which step a human still has to touch, and why. The grand plan assumed all of this away on a whiteboard. The small agent finds it out on Tuesday, for the price of a single tool call, while there is still time to be wrong cheaply.

This is the whole argument. Being wrong fast is recoverable. Being wrong slow, after four months of building on a guess, is the expensive kind. The small agent is not less ambitious. It is ambition that has agreed to be corrected.

There is a second payoff, quieter but bigger. The small agent earns the right to grow. Once it has tagged ten thousand tickets correctly, nobody argues about whether it can be trusted to do step two. It has a track record. You extend it because the evidence says you can, not because a slide deck said you should. The grand plan asks for trust up front, before it has done anything, which is exactly when trust is hardest to grant. The small agent buys trust one boring task at a time, and trust bought that way is the kind that holds.

The trap is that small does not feel like progress. A single tool that tags tickets does not look like the future of work. It looks like a script. So teams skip it, reach for the architecture diagram, and confuse the size of the plan with the size of the result. But the result of the grand plan, for months, is zero. The result of the small agent, by Friday, is a queue that triages itself. One of those is shipping. The other is a meeting.

Pick the agent that can be in production this week. Give it one tool and one job it cannot get badly wrong. Let it run against real data, in front of a real person, and watch what it teaches you. Then, and only then, add the second job, because now you know what the second job actually is. That is not starting small because you lack courage. It is starting small because it is the fastest path to something real, and real beats impressive every single time.

The company that wins is not the one with the most elaborate agent on the roadmap. It is the one whose smallest agent is already doing a job nobody has to do anymore.