Spanfold demo · Spend Guard · simulation

A stuck agent, with and without a cap

I run the same scripted agent twice. One run has no guard. The other uses my loop detector and spend cap. The agent gets stuck partway through, on purpose.

Without guard

Not run yet

Steps
0
Tokens
0
Dollars
$0.0000

    With guard

    Not run yet

    Steps
    0
    Tokens
    0
    Dollars
    $0.0000

      Known limit: this is a scripted simulation. No model is called and no real run was measured. The agent's token counts follow a fixed formula (input grows by 60 tokens a step; a stuck agent also re-sends each failed attempt). The guard only checks before a step, so the last step can push spend slightly past the cap; the table shows by how much.