Visual explainer

The Agent Router Architecture

How a small classifier in front of big models sends each request to its cheapest lane.

Sending every request to the giant model wastes time on easy ones.

Five requests queue up and even password help pays the giant's 75 units. Easy questions burn the same fortune as hard ones.

A Cheap Bouncer

A small classifier reads each request first for one unit of cost.

A tiny classifier names the kind and scores it for 1 unit before the giant wakes. Ten requests cost 10 units to sort, and most never go further.

Scores Become Addresses

Typed scores send each request to code, the giant, or a human.

Reset scores 0.90 and goes to code, essay scores 0.50 and goes to the giant, angry refund scores 0.80 and goes to a human. Bar length matches the score.

Mind the Gate

Scores below 0.8 skip the shortcut and go to a human check.

The 0.90 sails through while the 0.50 waits for human eyes. The gate at 0.8 splits the confident from the questionable every single time.

The Bill Shrinks

Routing cuts the bill from 750 units to 160 in this ten request sketch.

Ten requests at 75 units cost 750 before. After routing, 10 units of sorting plus 2 giant calls cost 160, about 4.7 times less. Units are illustrative, not a benchmark.

Set the Gate

Raise the gate when mistakes cost more than delays.

Run 0.8 for fast triage where errors are cheap. Crank it to 0.95 for refunds and safety, and drop it when your humans drown in reviews.

Where It Breaks

A confident wrong score sends the refund down the wrong lane.

An angry refund scored 0.91 as code skips its human and auto fires. Overconfident scores dodge the gate, so audit your low margin calls.

The Quick Version

Every request paying 75 units wastes easy ones.
A 1 unit classifier sorts before giants wake.
Scores 0.90, 0.50, 0.80 pick three lanes.
Below 0.8 waits for human eyes.
Ten requests fall from 750 to 160 units.
High stakes raise the gate toward 0.95.
A 0.91 mistake auto fires the wrong lane.