Chapter 12 · Blog
Tutorials walk you from npm install to a deployed
pipeline. Essays cover the mental model — autonomy ladders,
trigger patterns, why most "AI agents" don't survive their
first month in production.
Step-by-step: 6-agent democratic swarm with three rounds, evidence-weighted aggregation, verifier veto. Watch a 3-to-2 Critical majority lose to two voters with stronger evidence.
Majority alone does not decide truth. Evidence decides truth. Policy decides permission. Humans approve risk. The architecture /ll-swarm-vote scaffolds — independent answers, peer review, evidence-weighted aggregation, verifier veto.
Step-by-step: /ll-swarm-supervisor scaffolds the 4-agent OpenSyber MVP from the swarm blog §13. Tool permissions per agent, shared state in Postgres, executor gated by approvals inbox, every step audited. End-to-end in pipes.
The mental model. Workflow as PM, supervisor as team-lead, specialists as workers, verifier as QA, audit log as compliance memory. Five patterns, ten rules, and the OpenSyber-shaped swarm Luna's /ll-swarm-supervisor scaffolds.
Step-by-step: scaffold an agent with /ll-agent-build, expose it as a Luna verb with /ll-agent-call, race five variants with /ll-agent-swarm, audit with /ll-no-bluf, and ship to Cloudflare. ~15 minutes end-to-end.
Production AI agents need eight components and most teams ship two of them. /ll-agent-build scaffolds all eight — planner, RAG, MCP tools, verifier, guardrails, approvals, audit, eval — in one command.
The mental model. When to use an agent. When not to. Autonomy levels. Trigger patterns. The architecture that /ll-agent-build implements.
Posts are kept as Markdown in blog/ on
GitHub
so they ship with the npm package and stay grep-able from
/ll-rag.