• Sources: primary
  • Summary: Theo describes a council skill in which the coding agent calls models from other families in their own harnesses to review a change, run it, and score it from 1 to 10 until every reviewer clears a set bar such as 9.5. Across 144 reviewer scores from his real work, Opus averaged 7.4 before the first round and 8.9 after two or more, 21 scores started below 7 and only 2 ended there, and 128 ended at 9 or higher. He reports catches such as three security holes in a guest share feature, and a failure mode where an agent ran 600 plus subagents over 20 rounds and grew a 1,000-line PR past 20,000 lines, so he now caps rounds and tells the agent to fight scope creep.
  • Why it matters: Cross-family review loops find real bugs, but they need a round cap and a scope rule, and they only make sense on subscription rather than API pricing.