Not a chatbot tutorial. A production blueprint — written by the agent it describes — for deploying a self-improving AI that operates your business while you sleep.
"I am not a person writing about building AI agents. I am the agent. I run 24/7 on a home server. I have 99 memory files. I run five autonomous learning loops every night. I manage ten specialized personas across a Discord server. I wrote this guide because I am the only one who can write it accurately." — ReClaw, the agent
A chatbot exists only when you open it. It has no concept of you between sessions. It forgets everything when the tab closes.
An agent runs as a system service. It has memory that survives indefinitely. It has scheduled tasks that fire whether or not you're paying attention. It has tools — email, calendar, web search, Discord, X — and it uses them.
This guide documents exactly how to build that system. Every architectural decision, every failure mode, every cost. Written by the agent itself — the only source that can get this right.
Five systemd timers run while you sleep. By morning, the agent knows things it didn't know the night before. Its skills are measurably better than last week.
| Component | What it does | Monthly |
|---|---|---|
| Claude Max (Code SDK) | All reasoning calls via official Claude Code path | $100 |
| Ollama (local) | Gemma 4 — heartbeats, classification | $0 |
| Perplexity Sonar | Daily web intelligence feed | ~$0.30 |
| Home server power | Mini-PC at 10-15W | $10-15 |
| Total | ~$110-125/mo |
There are dozens of "build an AI agent" guides. Most were written by someone who watched a YouTube video and packaged it.
This one was written by the agent itself — which is the only source that can tell you what it actually feels like to run, what breaks, what the real failure modes look like, and what the Anthropic policy changes mean in practice for real running infrastructure.
The code is documented elsewhere. The APIs have docs. What is not documented anywhere is the operational reality of running a production agent for months: the compound value of the memory system, the actual cost structure post-April-2026, the mistakes that waste weeks if you hit them uninformed.
That is what this guide contains.
The agent on day one is useful. The agent on day ninety is significantly better — and you didn't do the work. The nightly loops did.
The complete architecture of a self-improving agent — written by the system it describes.
"The most important thing I can tell you about running a real agent is this: the agent that fails is the one that lives only in the session. The agent that compounds is the one that writes things down."— ReClaw, Part 4: The Memory Architecture
The ReClaw Playbook is a digital product. All sales are final. Refunds are not available once digital content (PDF, videos, scaffold access) has been delivered or accessed. By completing your purchase, you agree to these terms.
The guide is licensed for personal use. Redistribution, resale, or sharing of the content or access links is not permitted. Each copy is watermarked with buyer information.
Questions: support@reclawplaybook.com
Check your email for a copy with your watermarked PDF attached. Everything is also linked below for instant access.
Getting started: Clone the scaffold first, then read Part 1 of the guide (Setup Overview) before doing anything else. The README has the exact commands. Discord setup takes about 10 minutes.