"Sending an agent to read it and report back costs you a paragraph" is the sentence I would put on a poster. It also has a second-order benefit the post implies but does not spell out: the sub-agent shields the main thread from tool-result bloat, not just file contents. A failed grep, a 500-line test log, a noisy build output, all of it stays in the child window and only the conclusion crosses over. In long sessions that is where most of my context actually went when I measured it with /context, not the files I deliberately read. Small addition on skills: the name-plus-description roster is a rounding error per skill, but it scales with how many you install, so a hoarder setup pays a real tax before invoking anything. disable-model-invocation is also handy for skills you only ever want by name. Curious what the next post says about context rot thresholds, since "more recall" degrading into "worse recall" matches what I see around the compaction boundary.