AI agent design principles work like a well run office: every rule is written down once, everyone knows their job, and nobody ships work they checked themselves. Daniel Agrici showed this in a video on Oct 26, 2026, touring a vault of seventeen marketing agents and closing with ten rules on one screen. Below are those ten ai agent design principles, word for word, with what each one fixes and the one thing we would add.
The ten ai agent design principles, and what each one fixes
1. “Link, don’t copy. Each fact lives in one place.” Copied rules drift. When the brand voice lives in nine prompts, it changes in three and stays stale in six.
2. “Separate who an agent is from what it does.” A role note and a soul note. You can move the writer to email without losing how she writes.
3. “Keep agent files small. Push shared rules down into the team layer.” A short agent file is one a person can actually review.
4. “Give each agent its own expertise and shared rules from one source.” Deep knowledge per role, common rules for all. The blog agent does not need the accountant’s brain.
5. “Write procedures down once and reuse them.” Starting a session, reviewing work, asking for approval: write each once as an SOP.
6. “Give every input and output a template.” Templates make work easy to compare and easy to check.
7. “Set limits before adding power.” Every agent gets a ceiling, an off switch and a clear list of what needs a human yes.
8. “Check work with someone other than the author.” A reviewer or verifier that did not write the draft.
9. “Turn every mistake into a rule, a test or a template change.” That is how the system improves instead of repeating itself.
10. “Keep readiness separate: a task can be ready, an agent can be ready, and permission to run is a third, separate gate.” Ready is not the same as allowed.
Why these rules sound familiar
I’ve built, scaled, and sold three companies, generated over $85 million in the last seven years in my own companies. Every one of these ten rules is something we learned the hard way with people first: one source of truth, job descriptions, checklists, sign off limits, a second pair of eyes, and a post mortem that changes the process. Agents just make the cost of skipping them show up faster.
How to design AI agents with these rules this week
Do not try to apply all ten ai agent design principles at once. Start with principles 1, 7 and 8. Move every repeated rule into one shared note and link to it. Give each agent a written ceiling (draft only, no sending, no spending). Then name a reviewer for every output, human or agent, that is not the author. Those three alone stop most of the messes we see: drifting voice, agents that act without asking, and errors nobody catches.
What we would add to the ai agent design principles
His list is strong on structure. Two gaps we would close for a small business:
- A cost line. Each agent should say what it may spend in tokens or tools per run, next to its ceiling.
- A plain language owner. Each agent should name the person accountable for its output, not just a reviewer agent.
Both fit inside his limits layer, so they do not break the model. The full method, layer by layer, is in our guide to ai agent architecture.
Where he is right, and where to be careful
He is right that one giant prompt does not scale, and that writing the same rule in many places is the main source of drift. Be careful with one claim: he says skills are barely needed once the folders are right. His own blog agent links 33 skills. Read that as “skills are one layer”, not “skip skills”. These are also ai agent best practices for marketing work, not proof from a controlled test; treat them as a sound starting design.
Want agents built on these rules?
We build AI agent teams for service businesses with these principles baked in: one rule book, role and voice notes, written limits, an independent reviewer and a learning log. We also train your team to run and extend it. See the AI automation agency service and book a call to map your first agent.
Sources
- Daniel Agrici, “I’m deleting this video in 48 hours”, YouTube, Oct 26, 2026: https://youtu.be/aOI6VJzvqbE
- Founder figure: Tom Patterson’s own companies, seven years to Oct 2026
Would your agents pass a new hire review?
If you would not let a new hire work without a job description, a checklist and a manager, do not let an agent. Book a call and we will review one of yours.