Run the queue
Ninety seconds of inbound work, four places to send it, and a budget that runs out. Every route is a decision about what an agent is allowed to handle, which is the decision that actually determines whether an AI system holds up in production.
Inbound
Finds and quotes. Cheap, fast, narrow.
Reads context and writes. Costs real money.
Touches other systems. Gated for a reason.
Slow, expensive, never wrong.
Ninety seconds of inbound. Every route costs money and every promise has a clock. The cheap agent is cheap because it only does one thing.
What you were actually doing
Routing is the product
The models in the sim are fixed. The only thing you controlled was which work reached which one, and that alone decided whether the round was affordable. In a real deployment this is where most of the engineering goes, and it is the part that never appears in a demo.
A narrow agent is a feature
The retrieval lane resolves one kind of work for a dollar. It looks weak next to a lane that handles everything, until the volume arrives and the cheap lane is the only reason the budget survives. Narrow is what makes an agent safe to trust and cheap to run.
Rework is the hidden cost
A ticket sent to a lane that cannot resolve it does not fail loudly. It comes back later, older, having already been paid for. Systems without an evaluation loop do this continuously and the invoice is the first place anyone notices.
Somebody has to be the human
The unclear tickets cannot be resolved by any agent in the sim, and the only winning move is to pay for a person. Designing that route deliberately, rather than discovering it in an incident, is most of what a handover is worth.
This is the shape of the work we do: deciding what each part of a system is allowed to touch, and proving it holds before it reaches your customers.