Jun 2026
What an agent costs
Alex Research Economics
We expected the intelligence to be the expensive part. It is now close to free — which changes what a company is actually selling.
The assumption everyone started with was that tokens would be the cost centre — that running agents at any real volume would be the thing that broke the model. It went the other way. The same work costs a fraction of what it did a year ago, and keeps falling.
When the intelligence approaches free, nobody is buying intelligence any more. They are buying the ability to hand something over and be able to answer for it afterwards.
Every run is on the record
Usage is recorded server-side, per agent, per run — what ran, how long it took, and what it consumed. It appears on the agent's own passport, next to the work it did.
This is not analytics for its own sake. An agent you cannot cost is an agent you cannot manage, and 'it used some credits somewhere' is not an answer a business can plan around.
Priced like a worker, not like a seat
Per-seat pricing was invented for software that sits still. Agents do not: one may run all week and another twice a month, and charging the same for both is a fiction everyone can see.
So the meter follows the work. You pay for agents that actually do something, which is also the only pricing that survives contact with a sceptical owner.
"When intelligence is nearly free, the scarce input is not the model. It is trust."
What we do with the surplus
Cheap inference does not mean careless inference. It means the budget moves to the parts that were always harder: identity, permissions, memory, and proof — the work that makes an agent safe to leave alone.
That is where the cost went, and that is what the passport is.
Alex Research writes product notes for Rivesa, focusing on agent economics, metering, and the craft of keeping delegated work accountable.
