What a gauntlet run actually costs

The generated prompt tells your agent to keep going until you stop it. That is deliberate — a round count is a finish line, and a finish line is what stops work before the bar is met. But it has a consequence people meet the hard way: the run has no natural end, so it has no natural cost either.

This is not a warning against running it. It is the arithmetic, so you can decide what you are buying before you start.

Where the tokens actually go

Most people assume the builder is the expensive part. It usually is not.

A single round on one piece is: the builder produces something, then a fresh critic reads the real artifact and judges it. The critic starts with an empty context window and has to load whatever it needs to form an opinion — the rendered page, the file, the test output, the running app. That loading is input tokens, and it happens again on every round, because the critic is fresh every time.

Freshness is not an implementation detail you can economise away. A critic that remembers the previous round inherits its excuses; it starts agreeing with reasoning it has already seen. The cost of an independent judgement is paying for the context every time.

So the shape is roughly: one build, plus one full context load per critic per round, times however many rounds it takes. The builder's output is the small number in that expression.

Why Graph mode multiplies instead of adds

In Loop mode, a piece that takes five rounds costs about five builder-and-critic pairs. Bad enough, predictable.

Graph mode changes the exponent. A piece is only done when it traverses its whole graph clean in one pass, and any failure sends it back to the entrance. A piece that fails its fifth trial does not re-run one trial — it re-runs all five.

The rule that makes it worse is the good one: after every clean traversal a graph engineer adds a harder node. So the graph a piece faces on its eighth attempt is longer than the graph it faced on its first, and every reset re-runs the longer version.

There is no bug here. This is the method working — it is what stops the builder learning the shape of a fixed test suite. But it means "Graph mode costs somewhat more than Loop mode" is the wrong mental model. Budget for a multiple, not a margin.

Three habits that keep it sane

Stop pieces, not runs. The graph grows a node after every survival, so a piece can improve indefinitely. There is no point at which it declares itself done. Watch the progress page and retire each piece as it clears the bar you actually have, rather than waiting for a run that will not end on its own.

Give the bar something to hit. The single largest cost driver is a vague bar. "High quality" gives the critic nothing to check against, so it invents a standard, and a piece can fail an invented standard forever. "As legible as a train departure board" or "passes the suite in /tests" terminates. If you supply no bar the generator picks one for you and names it in the first sentence — read that sentence, because it is what your money will be spent pursuing.

Split before you run, not during. The lead agent decomposes the goal, and how it splits determines how much re-running a failure causes. Pieces that overlap mean one fix breaks a neighbour's approved trial, and both reset. You cannot control the split directly — telling the agent how to decompose is exactly the kind of implementation prescription the method removes — but you can control how ambiguous the goal is, and an ambiguous goal produces overlapping pieces.

What this tool costs, for contrast

Generating a prompt here is one short model call: roughly a cent, once. That is the whole transaction on our side, and it is why the daily allowance is three rather than three hundred — not because generation is expensive, but because a prompt you did not think about is a run you will not watch.

The run is the expense, it happens on your account, and nothing in the prompt will slow it down for you. That is the trade the method makes: it removes every stopping rule except your judgement, and hands you the bill for the ones you do not exercise.