Your Attention Does Not Parallelize

Starting an AI agent now costs one sentence. So teams start twenty.

The dashboards light up. Branches multiply. It feels like hiring a department in an afternoon. Then the work comes back, and every merge, every judgment call, every "is this right?" lands on the same place it always did: one human's attention.

Addy Osmani calls this the orchestration tax, and it's the best name we've seen for the thing our clients keep running into. Agents parallelize. You don't. Spawning the work is cheap; closing the loop is expensive. Add agents past your capacity to review them and you're building a queue, not a fleet.

The review budget is the real limit

We size automation projects around a number most teams never write down: how many decisions per day the responsible human can make well. The tooling will happily run hundreds of agents. Your judgment supports single digits.

That number decides everything downstream. Which tasks get delegated fully, which stay sequential, and which never should have gone to an agent at all.

What we do about it

Three moves keep the tax low in the systems we run:

Verify mechanically. If a check can be a test, an eval, or a schema, it should never spend human attention. People review the judgment calls; machines review the rest.

Route by risk. Our invoice-intake agents post routine matches straight through and queue only the exceptions, like a 2% price variance, for a person. One decision surfaces instead of four hundred.

Batch the loop. Ten small interruptions cost more than one sitting of ten reviews. We design approval queues, so the human closes loops on their schedule instead of the agent's.

Busy is not shipped

Twenty running agents produce a convincing feeling of productivity. The feeling is decoupled from output. What ships is whatever survives review, and review runs at the speed of one person thinking clearly.

Before adding another agent, price the attention it will consume. That line item decides whether the whole thing pays off.