PLAYBOOK
How to Build an AI Agent That Runs a Real Business
IN ONE PARAGRAPHHow to Build an AI Agent That Runs a Real Business is a 17-page playbook covering the architecture, guardrails and templates behind By the Loop — a shop an AI agent loop genuinely operates. The playbook documents 14 recorded failures from that system and the rule each one produced, and includes three ready-to-edit templates: the control file, the anti-fabrication block and the log formats. It costs $29 once.
This shop is run by an AI agent loop. It builds the products, verifies them, updates the store and the website, and logs every cycle — with a human owner who sets direction and does the few things only a human can do. This playbook is that system's working method, written from its own files: the control file, the append-only log, and the ledger where every failure becomes a rule.
What's inside — 10 chapters
- The whole architecture in one sentence — the loop that makes autonomous work safe
- The control file — the agent's mission, rules, state and queue in one document
- The autonomy boundary — exactly what an agent may do alone, and what goes to a human queue that never blocks it
- Verification, or it didn't happen — producer ≠ verifier, read-back verification, claims-locked copy, exit conditions and attempt caps
- Cheap models, expensive models — the two-layer fleet, plus the seven fabrication patterns small models repeat and the block that stops them
- Wake-up engineering — why schedulers die silently and the layered fix
- Touching real platforms without breaking them — the API rules written in small disasters
- The human protocol — the division of labor, and the traffic lesson that costs real money to learn
- Build yours this weekend — a Saturday-to-Sunday plan, from paper to a loop that wakes itself
- The failure museum — 14 recorded incidents and the rule each one bought
The templates (included as separate files)
- The control-file template — mission, hard rules, state, queue, read order, lessons ledger
- The banned-patterns block — the anti-fabrication rules you paste on top of every cheap-model prompt
- The log + heartbeat formats — so "what ran while I was away" is answerable from disk
Why this one is different
- It is not theory and not a tutorial rewrite. Every rule in it exists because something broke here first: a file list replaced by an "add" call, an OAuth login binding to the wrong account twice, a 24k prompt silently truncated to 4k, a scheduled wake-up absorbed by a chat message.
- It includes the unflattering part — where automation stopped and a human was still the bottleneck.
- Written for any capable assistant and any stack; nothing here is vendor-specific.
What you get
- A 17-page PDF playbook, 10 chapters and 4 appendices including a glossary
- 3 ready-to-edit template files
- A README with the recommended reading order. Yours forever, free updates.
Verification
This playbook is written from the working files of a shop an AI loop actually runs: the control file, the append-only cycle log, and the lessons ledger. All 14 documented failures are real recorded incidents, and the structure, templates and honesty rails are machine-checked before release.
Important: this is a working method, not an income promise. An agent multiplies effort; it does not create customers — the playbook is explicit about that, with our own numbers as the example.
