5 lessons · ~40 min · free
Run AI agents that don't lie, leak, or overspend
This is the field manual for a live, self-hosted agent fleet that does real, load-bearing work: discovery agents that read the open web, authoring agents that draft claims, an independent verifier, a human-gated publishing lane. Every rule here is the scar tissue of a real failure (a stuck pipeline, a leaked credential, a prompt-injected web page, a silent overnight spend). Built by Paul Takisaki.
Not a framework, not a tutorial for any one tool. The patterns are the point. Start at Lesson 1 or jump in anywhere.
-
1
Why Agents Fail
Self-review is correlated with a model's own blind spots. Why structure beats prompts, and what a real fleet looks like.
-
2
The Verification Wall
Create is never verify: the author of a claim can never approve it. The wall lives in database grants, not prompts.
-
3
The Money Lesson
The cheapest model that clears the bar, empty fallbacks, and governors that refuse to spend blind. $50/day to $0-2/day.
-
4
The Hostile Web
Least privilege, prompt injection, and secrets: the web is data, never instructions.
-
5
Run It Like Production
Live config is truth, one change at a time, and the dispatch sequence that never double-spawns.
All five lessons are live. Work through them in order, or jump to what you need.