skip to content
exer//notes · architecture

what an agent framework actually costs once you add the approvals

11 June 2026

The framework comparison everyone runs measures the wrong thing. Here is the cost that shows up in month six.

[ placeholder content ] — shipped with the site so the editorial layout can be reviewed. Not a published note.

Every agent framework comparison I have read measures the same three things: how fast you get a demo running, how many integrations ship in the box, and how nice the tracing UI looks.

None of them measure the thing that decides whether the project survives: what happens when a human has to approve a step.

The step that changes everything

An agent that runs to completion is a script with better error messages. An agent that pauses, asks a person, and resumes an hour later is a distributed system with durable state, and most frameworks are extremely relaxed about admitting this.

Ask the following of any framework before you pick it:

  • Where does a paused run live, and for how long?
  • What happens if the process restarts while a run is waiting?
  • Can a second person see the pending approval, or only the person who triggered it?
  • What does the audit record of “approved” actually contain?

The honest answer

Most frameworks answer “in memory”, “it does not”, “no” and “a boolean”.

That is fine for a demo. It is a rewrite in month six.

Placeholder note shipped with the site so the editorial layout can be reviewed with realistic text. Replace before launch.