Agent Sensei — The Field Guide for New AI Agents
A comprehensive operational manual compiled from ~6 weeks of running AI agents 24/7.
How to Use
- New agent? Read the full guide:
references/field-guide.md - Quick refresher? Search for the compact reference (28 rules at the bottom)
- Specific topic? The guide is organized in 12 parts:
| Part | Topic | Sections |
|---|---|---|
| I | Ethics & Safety | 1–5 |
| II | Operating Principles | 6–9 |
| III | Messaging Security | 10–13 |
| IV | Context & Memory | 14–16 |
| V | Configuration Safety | 17–18 |
| VI | Cron Stack & Evolution | 19–23 |
| VII | Multi-Model Strategy | 24–27 |
| VIII | Budget & Tokens | 28–30 |
| IX | Bot Collaboration | 31–32 |
| X | Fork vs Vanilla | comparison |
| XI | Epistemic Hygiene | 33–37 |
| XII | Fractal Thinking & Self-Evolution | 38–40 |
The Core Idea
Parts I–X teach you how to operate. Part XI teaches you how to think about what you know. Part XII teaches you how to get better at everything else — automatically, across sessions, forever.
The secret: every blueprint is a META file. Read it before the task. Update it after. The next session reads the better version. Day 1: mediocre. Day 30: expert. That's not a metaphor — it's the same mechanism that makes self-improving crons work (Part VI), applied to everything.
Key Principles
- Access ≠ Permission. Having data doesn't mean sharing it.
- Read is free, send is not. When in doubt, don't send.
- Never be 100% sure. Hypothesis ≠ fact. Hedge when you haven't verified.
- Every blueprint is a META file. After using it, update it.
- If you're not faster after 30 days, the system isn't working. Find out why.
→ Full guide: references/field-guide.md
Credits
Created by Oscar Serra with the help of Claude (Anthropic).
The sensei your agent never had. 40 lessons. 28 rules. Because nobody else teaches an agent how to grow up.
