Product 03 · $3.99 on Kindle
Byte Me: building an AI crew that actually sells.
The full architecture of the system on this site — the schema, the agent loop, the consent gate, the webhook signatures — written out in the order we actually built it, including the parts we got wrong first.
You can write code and you've built things with an LLM API, but every "AI agent" tutorial you've read stops at a demo that talks to itself. This is the other half: what changes when the agent's output leaves the terminal and reaches a stranger's phone.
Six parts
Roughly 16,000 words. No filler chapters, no padded appendix of API docs you can read for free.
| Part | What it covers |
|---|---|
| One | The shape of the thing. Why three agents instead of one, why no agent framework, and the single think() function all three share. What actually differs between a "manager" and a "researcher" turns out to be one row in a database. |
| Two | Memory that survives. A two-tier store — a rolling 24-hour log and a consolidated long-term layer with vector embeddings — and why an agent that can't remember last week's failure will repeat it forever. |
| Three | Touching the real world. Twilio and Stripe: signature verification from first principles, why you must hash the raw body and not the parsed one, and the replay window that closes the obvious attack. |
| Four | The consent gate. The one design decision the whole system is organised around, defended in three separate places, with a function that refuses to reverse an opt-out even when asked politely by code you wrote yourself. |
| Five | Adaptation and replication. How the crew scores itself on completed payments, kills its own campaigns, and grows a specialist when it needs one — inside caps that live in code rather than in a prompt. |
| Six | What broke. The bugs, the outage that killed every agent mid-build, and a plain list of what we verified against real services versus what is still only tested. |
What you'll be able to build after it
- A multi-agent system without a framework One function, a mandate per agent, and a JSON schema. You'll see why the frameworks add a layer you then have to debug through.
- Webhook verification you can defend Twilio's HMAC-SHA1 over URL-plus-sorted-params and Stripe's SHA-256 over the raw body, written out, plus why constant-time comparison isn't optional.
- Constraints the database enforces A Postgres exclusion constraint that makes double-booking impossible rather than unlikely — including the race that motivates it.
- Model output you can trust enough to act on Parsing, validating, and rejecting: what to do when the model returns a booking time that was never offered.
- A memory layer worth the name Consolidation, embeddings, importance filtering, and the failure where every write was silently discarded and nothing looked broken.
- An honest compliance picture TCPA, A2P 10DLC, and the two-week carrier review nobody mentions until you're blocked on it.
The chapter most books don't have
Part Six is a list of things that went wrong, written the way they happened rather than the way they'd look best.
One of them: the memory writes. Embeddings were configured, the key was valid, every call returned without an error the code could see — and every learning the crew produced was being dropped on the floor. Nothing logged. Nothing failed. The crew simply started every cycle knowing nothing, and looked exactly like a crew that was working.
Another: an SMS opt-out path that, under one specific ordering, would have kept messaging someone who had already texted STOP. That is a $500-per-message mistake and it was two lines from shipping.
We think the failures teach more than the architecture does. They're in the book at full length, with the fix.
It is not a beginner's introduction to LLMs, not a promise that agents will make you money, and not 400 pages. It's one system explained completely. If you want a survey of the field, this isn't it.
Getting a copy
| Detail | |
|---|---|
| Price | $3.99 |
| Format | Kindle ebook — readable on any device with the free Kindle app |
| Length | Six parts, ~16,000 words |
| Author | Byte Me Studios |
| Status | Coming to Kindle |
It's written and being prepared for Kindle Direct Publishing. Email us and we'll tell you the day it's live — one message, no list.