For most of a year the plan was the obvious one: work on the product that needs it most, and let the others wait. It is a reasonable plan and it produces exactly one product's worth of progress, which is what it sounds like it would produce.
What changed is not that the agents got faster. It is that the cost of a second lane went to nearly nothing while the cost of reviewing a second lane stayed exactly where it was. Once that asymmetry is the thing you design around, the shape of the floor falls out of it: many lanes, one gate, and everything before the gate arranged to make the gate fast.
That means small branches. It means the spec is written before the code, because reading a diff against a stated intent takes a fraction of the time of reading a diff and inferring one. It means the adversary agent has already attacked the branch and attached what it found, so the review starts from what is known broken rather than from nothing.
The number that matters is how long a finished branch waits before it gets read. Ours is usually under a day. When that number goes up, we close a lane.