Your Attention Does Not Parallelize

Starting an AI agent can take one sentence. Reviewing its work takes longer.

The dashboards light up and branches multiply. Then the work comes back: merges, judgment calls, and changes that look plausible but need a closer read. They all compete for the same person’s attention.

Addy Osmani calls this the orchestration tax. His point is that agent output can grow faster than your capacity to review it. I think that’s a useful constraint to put at the start of an automation project.

Set a review budget

How many decisions can the responsible person review well in a day? How much context does each decision require? An invoice exception and an architectural change consume very different amounts of attention.

That budget helps decide which tasks can run independently, which should stay sequential, and which need the person involved from the beginning. The number of agents the tool can launch tells you very little about this.

Make the queue manageable

I’d start with three habits:

• Have the system run tests, schema checks, and evaluations before requesting review. Show those results alongside the work.
• Route decisions according to risk. An invoice workflow might process routine matches within agreed limits and flag a price variance for review.
• Collect related reviews into a queue so someone can work through them in one sitting, with the relevant context close at hand.

Measure completed, reviewed work alongside the backlog. If the queue keeps growing, reduce the incoming work or change what the system asks people to decide before adding another agent.