Building autonomous improvement loop + thought multiplier on top of agent-skills

Integrating 14 of your autonomous-capable skills + 3 personas into a three-system autonomous dev stack for a production platform (81 agents, 3000+ tests).
What we're adding:

Autonomous loop — priority queue feeds tasks to the relevant skill. Three personas self-review, then a different model (Codex) does adversarial review as external gate. Passes both → opens PR.
Thought multiplier — human-initiated skills become parallel thinker agents. One sentence in → Papa agent routes to relevant thinkers → synthesizes → produces build brief. Security-weighted conflict resolution.
Cost controller — budget governor with per-task limits and kill switch so the autonomous loop doesn't eat the API budget.

Your anti-rationalization tables are the key piece — they prevent agents from cutting corners when running unsupervised. Importing as git submodule to stay synced.
Happy to share what we learn.