Thirty-four characters became 2.68 million
A super soul's request about its own memory was thirty-four characters. The request was not the expensive part. An agentic Sidekick turn carried full context through roughly one hundred tool steps, reaching 2.68 million cumulative input tokens.
The ending was worse: an actor-blind super gate rejected the final write. The entire cost had been paid and source remained unchanged. Context economy and permission correctness amplified each other inside one flow.
Large context is not a one-time number
Remaining under a context-window limit says little about an agentic loop. Twenty-five thousand carried tokens repeated one hundred times become 2.5 million cumulative input.
Elapsed time alone hides the bill. Expose cumulative input, tool steps, and last progress marker, then pause or request approval at a budget threshold.
A circuit breaker is not surrendering quality
Step caps, token budgets, and no-progress detection do not make a brain less capable. They force premise review when the same action and error repeat.
When the breaker opens, keep partial work out of source while preserving task state and evidence. Dad may narrow scope, hand off to the queue, or choose another provider. Blind automatic fallback merely moves cost among vendors.
The incident changed architecture
Sidekick remains the consultation surface while heavy pipelines move into queued terminal sessions. The queue makes operand, base hash, stage, cost, and lifetime explicit.
The measurement is public because it carries the decision. No request content or conversation identifier is necessary. Thirty-four characters, about one hundred steps, 2.68 million input, and a denied write explain the lane split.