Michael Mignano made a useful point about the coming end of “tokenmaxxing.”
I think he is right.
The first answer will be model routing. Cheaper models for cheaper work. Expensive models where they actually change the outcome.
But the thing I keep coming back to is Google.
Some people are much better at Google than others.
Usually it is not because they know a magic operator. It is because they know what to ask.
They know the company changed its name before the lawsuit. They know the acquisition came after the product launch. They know the useful source is probably a filing, not a blog post. They know which journalist covered the first version of the story. They know the phrase insiders use, which is often not the phrase everyone else uses.
That is context.
Most of it is temporal context: a rough map of what happened, when it happened, and what came before what.
Humans take this for granted because we carry it around in our heads. Agents do not. Not yet.
When an agent goes to the open web, it often starts cold. It spends tokens doing the part a good human searcher has already done before typing the query.
It works out what changed, what is stale, which source matters, which result is an echo, and what question it should have asked in the first place.
That is where a lot of spend disappears.
At Synorb, our bet is that search should start after context, not before it.
Give the agent a temporal map first. Then let it search.
The result is not just cheaper search. It is better search.
Because the agent is no longer trying to discover the shape of the world from a ranked list of links. It arrives with enough context to ask a better question.