An AI agent runs an endless growth loop over your knowledge site — sensing its health, researching what matters, triaging issues, acting, creating and distributing content, then feeding every result back into the next cycle. There is no admin panel. There is a loop.
⚙ Phase 4 — MEASURE is live: per-article search metrics, page-2 opportunity mining, gated rewrites. Dashboard →
Each cycle is a resumable "tick" — not a runaway while(true).
A scheduler fires the tick; the agent walks the stages guided by playbooks; every result is measured and written back.
Analytics, Search Console, access logs, revenue and uptime become structured observations in the loop's memory.
Trends, sources and competitor signals are gathered by collector adapters — the raw material for decisions.
Observations become a prioritized issue queue: traffic drops, indexing errors, content gaps, stale pages.
The agent works issues via playbooks — SEO fixes and site changes land through git worktrees as pull requests.
Articles, videos and share cards are generated — and must pass quality gates (a separate verifier agent) before anything ships.
Publisher and announcer adapters push content to the site, social feeds and video platforms.
The agent is replaceable — any agent CLI (claude -p, codex exec) can execute a stage.
The loop is the product.
cron/systemd fires resumable ticks per cadence — sense hourly, create daily, act weekly. Crash-safe: every tick starts from persisted state.
Procedures are first-class files. The agent must read the playbook before executing a stage — no improvisation on production.
Pluggable sensors / collectors / generators / publishers / announcers. Swap Search Console for Plausible, FTP for GitHub Pages.
Creator and verifier are separate agents. Content that fails factuality or SEO gates is repaired or discarded — never silently published.
SQLite holds the issue queue, content ledger, metrics time series and a decision log you can audit.
Site-modifying actions run in isolated git worktrees and land as reviewable pull requests.
site: aiknowledgecms.exbridge.jp kpi: [pv_weekly, affiliate_clicks, indexed_pages] cadence: sense: hourly create: daily act: weekly budgets: articles_per_day: 3 llm_cost_per_tick_usd: 2.0 gates: [factuality_check, seo_check] escalate_when: [revenue_drop_30pct, index_errors] adapters: sensors: [gsc, simpletrack] publisher: ftp_heteml announcer: aixsns
The whole loop — KPIs, cadence, budgets, gates, escalation — is declared in a single file. Point the runner at a Loopfile and the site starts growing.
An autonomous infinite loop without guardrails is a liability. These are built in, not bolted on.
Hard caps per tick: LLM cost, article count, API calls. The loop stops before your wallet does.
Every stage can run with full reasoning and zero side effects. Audit before you trust.
One flag halts the loop. State is persisted, so resuming is safe.
Declared conditions page a human (email / LINE). Verification stays a human responsibility.
aiknowledgecms.exbridge.jp itself is being rebuilt as the framework's first instance: an AI-agent-economy media site grown by the loop — articles, videos, share cards and revenue, all produced and measured by the loop, with every tick's report published. The framework's proof is the site you are reading.
Manifesto, Loopfile specification, architecture. Public from day one.
Real sensors populate the issue queue in SQLite. Every tick publishes a public loop report — running hourly on this site.
The loop researches sources, generates articles, and publishes only what passes an independent verifier gate — rejected drafts stay in the ledger with reasons.
Live dashboard shipped — the loop publishes its own scoreboard hourly. Next: growth cards, a diagnosis experience, and site-modifying playbooks via worktrees.