Aug 23, 2026

A 4-minute essay on the brain refusing to parse documents that carry strong AI traces. Design docs with leftover Claude phrasing, decks mixing a sane strategy with generated filler: the author reads them without absorbing anything, then burns time in back-and-forth questions

Why it matters: It names the reviewer-side cost of the generation boom: cognitive wear. Together with AI;DR (covered Aug 18), the reader's burden is becoming this summer's recurring debate

An open-source multi-agent harness for running an office of your clones. It bundles 12 CLI agents (Claude Code, Codex, Gemini CLI and more) under your existing subscriptions' hourly limits, monitored through a deterministic Office-style simulation view

Why it matters: The Office homage supplies the charm while the economics supply the point: no extra API billing, just your idle subscription hours. It lands on a real problem, herding today's sprawl of CLI agents

Responding to reports that high effort felt nerfed, Anthropic's Thariq explained the team tests API serving configs in Claude Code before rollout, and one live config maps the numerical effort value differently. The number alone is meaningless; the effort you select is what you get, he said

Why it matters: A rare official confirmation that delivery-side experiments sit behind the model got worse feeling. Harness users just gained one more variable to check when quality seems to drift