How to Build Multi-Agent AI Systems That Actually Work in Production | Tyler Fisk

Oct 16, 2025

Tyler Fisk, an AI practitioner and instructor, shares his journey of building a successful AI education business in just one year. He explains the importance of rigorous human-in-the-loop reviews before deploying AI agents and introduces his creation, Gigawatt, which develops and evaluates other agents. Tyler discusses meta-prompting, the need for multi-agent architectures, and strategies for optimizing performance. He emphasizes the cost benefits of AI automation versus human labor, all while crafting a real-time Apple customer service solution.

Ask episode

AI Snips

Chapters

Transcript

Episode notes

ADVICE

Start With Human-In-The-Loop

Start production with 100% human review and gradually increase autonomy to about 60–70% as confidence grows.
Never ship an agent without a human-in-the-loop checkpoint to catch edge cases and policy risks.

INSIGHT

Divide Agents Like Company Roles

Build multi-agent systems by separating roles like experts and communicators to mirror real teams.
Different agents allow tuned temperatures and system instructions for deterministic research versus creative responses.

INSIGHT

Use Meta-Prompting For Iteration

Meta-prompting an agent to self-review and iterate raises quality from mid to high scores.
Ask the agent to score itself, explain gaps, and produce an improved V2 to reduce hallucinations.

Get the Snipd Podcast app to discover more snips from this episode

Get the app

Tyler Fisk built a $1.6 million AI education business in one year.

Zero PhDs. Zero Silicon Valley pedigree. Just a systematic approach to building AI agents that actually work in production.

While everyone’s vibe coding in ChatGPT, Tyler’s teaching thousands of students to build multi-agent systems for real businesses. Hundreds of production deployments. Actual revenue.

Today he’s doing a live build: Taking Apple customer service from idea to working multi-agent system in under 90 minutes.

No theory. Pure execution.

----

Brought to you by:

* 1. Maven: Get $135 off Tyler’s course with my code AAKASHxMAVEN

* Vanta: Get $1,000 off AI security & compliance at vanta.com/acos

* Testkube:

* Kameleoon: Leading AI experimentation platform

* The AI Evals Course for PMs: Get $1155 off with code ‘ag-evals’

----

Key takeaways:

1. Stop Vibe Coding: Most teams write one prompt, test twice, ship to production, and hope for the best. Tyler's rule: "We would never put it into production without a human-in-the-loop checkpoint. That's irresponsible." Start with 100% human review, gradually move to 60-70% autonomy.

2. Use Meta-Prompting to Build Agents: Tyler built Gigawatt—an agent with 72,000 characters of system instructions that builds other agents. It researches the domain, writes V1 instructions, evaluates itself (scores out of 100), identifies gaps, and rewrites to V2. Goes from 77% to 86%+ quality.

3. Build Multi-Agent Architectures: Don't build one agent that does everything. Separate concerns like you'd separate teams. For Apple: Core (expert agent, temp=0, finds facts) + Echo (email agent, temp=0.7, writes responses). Each optimized for its specific role.

4. System Instructions Need 7K-9K Tokens: Structure includes Role (job description), Context (business details), Instructions (step-by-step process), Criteria (guardrails), Examples (meta reasoning). Most people write 200 tokens. Tyler writes 7,000-9,000. That's the foundation.

5. Temperature Is Your Secret Weapon: Tyler's Toy Story analogy: Imagine an icy peak in a claw machine. Temp=0 (frozen): claw picks from top only—deterministic, precise. Temp=1 (melted): claw grabs anywhere—creative, varied. Match temperature to agent's job.

6. Information Hierarchy Prevents Hallucinations: Priority order: RAG database first (scraped company docs), System instructions second (built-in expertise), Web search third (with chain-of-verification). When agents search without verification, they hallucinate.

7. Build Complete Workflows: Tyler's 9-step production workflow with 5+ agents: Email arrives → Sentiment analysis (Cinnamon) → Expert research (Core) → Email writing (Echo) → QA loop → Human checkpoint (Slack) → Generative filter → Send → Log to memory.

8. Observational Evals Come First: Test 20+ different scenarios manually. Include edge cases and adversarial inputs. Document every failure. Save golden examples. Only after building confidence do you add systematic evals in production.

9. Calculate ROI as Labor Cost Reduction: Traditional cost: $460/day (expert time + customer service rep + manager review) = $138K/year. AI cost: $153/day (platform fees + API credits + human review) = $45.9K/year. Savings: $92K annual (67% reduction).

10. Emotion Prompting Actually Works: Tyler ends every prompt with "Go get 'em slugger." Based on research: positive reinforcement improves LLM outputs by ~15%. The same psychology that works on humans works on LLMs. "Be nice to your AI. They're gonna have robot bodies soon."

----