Show Notes
Mike Vitez has been shipping software since before the word “agentic” meant anything, and a few months ago he did something most experienced technical leaders refuse to do: he went back to the drawing board and picked the architect's pencil back up.
Funding stage: Bootstrapped. Saturnia Design is a founder-led UX and product studio Mike co-founded with Reka Szijj, growing on client revenue with no announced outside capital.
Ten years leading projects. A computer science background. Fifty plus products across three continents, by his studio's count. “I haven't been coding for a while,” he says on this episode. “Like, I wasn't the one who was writing the code, but my people. I needed a mindset change as well and go back to the real architecture part stuff. Not just reviewing it, but planning it.” That decision, made in a moment of frustration with routines that had gone stale, is the hinge this entire conversation swings on. Mike is now a member of the Claude Partner Network, running an in-house runtime environment where AI agents handle testing, development tasks, and product planning ahead of the launch of Saturnia Labs. He has automated quality control on every single commit. He has run forty-plus agents in parallel to build a product landscape in days instead of months. He is, by any reasonable measure, all the way in.
And his central warning is this: AI can save you the cost of a low-level code developer. It cannot save you the cost of an architect. “You can build a house without a foundation,” Mike says, “but it won't be solid.” That is not a caution against speed. Mike is aggressively pro-speed. His whole thesis is that MVPs are now trivially fast and you should rush to market, get users, and iterate. The caution is against speed without structure, which is a different animal entirely. It does not prevent collapse. It schedules it.
Ryan pushes on the obvious founder question: if you use Claude and I use Claude, what exactly am I paying you for? Mike's answer is one of the cleanest articulations of the AI-era consulting value proposition the show has had. Go to Claude and say “build me something,” and it will build you something. It might even work. But can you find the bugs? Can you make it scalable? Can you still add the fourteenth feature without the whole thing coming apart in your hands? “Most of the people cannot,” he says, “because if you don't plan it out right, that's where the real engineering knowledge comes into the picture.”
Mike does not treat models as tools. He treats them as a portfolio to be routed against. He is a Claude partner, and he also thinks reflexively defaulting to the most capable model available is a budget-burning mistake: “It's too robust for most of the tasks, and just burning a lot of tokens and spending a lot of money on things that shouldn't be necessary to spend money for.” When Kimi K3 landed with strong agentic benchmarks, his reaction was not loyalty. It was architecture: build a modular enough layer that a new model plugs in and gets tested on Tuesday.
And when Ryan raises the thing every founder building with AI has felt in their bones, that suspicion that the business model is quietly disincentivized from letting you finish, that the memory conveniently evaporates as you near the finish line, Mike does not dismiss it. “This is a super real problem,” he says. “And I think a lot of people are facing that.” His answer is not to complain about it. It is to build an environment where it cannot happen to you. That is the whole episode in one move: stop being a user of AI, start being the architect of the system that uses it.
Named Frameworks
The Foundation Test
Mike's core diagnostic for any AI-built system. Speed is not the variable that matters. Survivability is.
- →Can you debug it in six months when real customers depend on it?
- →Can you add feature fifteen without touching fourteen files?
- →Did anyone define the end goal before the first prompt, or did you discover it by prompting?
- →Would a new engineer understand the structure without you narrating it?
- →If the answer to any of these is no, you have a demo that's wearing a product costume.
Model Portfolio Architecture
The anti-vendor-lock layer. Don't adopt a model. Build the routing layer that models plug into.
- →Route each task class to the cheapest model that clears the quality bar, not the smartest model available.
- →Reserve frontier-tier models for genuinely hard reasoning, not for tasks a mid-tier model handles fine.
- →Run models in parallel for different job types inside the same workflow.
- →Keep the swap cost near zero so a new release becomes a one-day test, not a one-quarter migration.
The Commit Gauntlet
Mike's automated quality pipeline, and the automation he says he'd miss most if it vanished tomorrow.
- →A new commit lands from an agent.
- →An automated quality checker reviews it and can reject it back to the agent.
- →Only after the automated check passes does a human review happen.
- →On human approval, the commit and deploy execute automatically, DNS configuration included.
The Adversarial Agent Guardrail
Mike's mental model for why agent hallucination is a system design problem, not a model defect.
- →Treat agents the way you treat people: capable, useful, occasionally confidently wrong.
- →One agent produces. A second agent challenges the output.
- →The challenge is what triggers the rethink, exactly as it would with a colleague.
- →Reframe from "the model hallucinates" to "my architecture has no dissent in it."
Build With, Not For
The partnership motion Mike is actively running instead of traditional agency billing.
- →Find operators who already have a market, a client base, and a proven method.
- →Codify their method into a product instead of quoting them for a build.
- →Skip the invoice on the first pass. Test it with their existing clients together.
- →If it works with real users, formalize into a product with a profit share.
The Knowledge Base Guardrail Loop
The semi-automated customer support pattern Mike says most companies get wrong.
- →Ground the system in your own knowledge base, not a generic model's general knowledge.
- →Guardrail it so it declines rather than invents.
- →When it can't answer, escalate to a human admin instead of guessing, then absorb the answer into the knowledge base.
- →Mike's self-reported pilot result, three weeks in: roughly 700 questions handled, 83% answered automatically (one client, unverified).
Founder Experiment
Run the Six Month Debug Test, ninety minutes, on your most AI-built module. Pick the piece of your product that was built fastest with the most AI assistance, you know which one. Write down one small feature a real customer has actually asked for: one field, one filter, one export. Open a fresh session with no prior context loaded, no memory, no history, no you explaining the codebase, and ask for that one feature. Count two numbers: how many files had to change, and how many things broke that were previously working. Score yourself honestly. One to three files and nothing broke means your foundation holds. Four to eight files or one regression means you're carrying architectural debt with interest accruing. More than eight files or multiple regressions means you have a demo, and the collapse Mike describes isn't a risk, it's a date on a calendar you haven't looked at yet. Whatever the score, install one guardrail this week. Not a rewrite. One guardrail. An automated check on the next commit is the highest-leverage version.
Glossary
Agentic system
A coordinated architecture of AI agents that delegate, execute, and check work across a workflow, as distinct from a single chatbot answering a single prompt.
Vibe coding
Building by prompting your way forward without an architectural plan. Fast to a demo, expensive to a business.
Guardrails
Constraints, checks, and permission boundaries that keep autonomous agents inside acceptable behavior. Mike's position: without them the system falls apart.
Hallucination
When a model asserts something false with confidence. Mike frames this as a human-style error mode that should be caught the same way, with a second opinion.
Sub-agents
Specialized child agents spawned to work in parallel on pieces of a larger task. Mike reports using more than forty in parallel for one research project.
Vendor lock
Structural dependency on a single provider or model that makes switching costly. Mike treats avoiding this as an architecture requirement, not a procurement preference.
Modular architecture
A system built as swappable independent components, which is what makes model swaps and feature additions cheap instead of catastrophic.
FDE (Forward Deployed Engineer)
An engineer embedded directly inside a customer's environment to build custom solutions on site rather than shipping generic software.
Claude Partner Network
Anthropic's partner program, tied to the Anthropic Partner Academy, which gates access to certification exams and starts free at the Registered level.
Profit share
Mike's alternative to invoicing. Build the MVP without a bill, formalize revenue sharing if it works with real users.
Q&A: What Founders Ask After This Episode
Is vibe coding an MVP a bad idea for founders?
For a demo, no. For a business, it's a deferred cost. Mike Vitez, Co-Founder and CTO of Saturnia Design, frames it as building a house without a foundation: it stands until it's asked to hold weight. AI genuinely collapses the cost of low-level implementation. It does not collapse the cost of architecture. Vibe code the prototype, then decide deliberately whether the thing you're about to scale was designed or discovered.
Should a founder standardize on one AI model?
No. Mike, a Claude Partner Network member, argues that always reaching for the most capable model burns tokens on tasks that don't need it, and that committing to a single model is a vendor lock risk. His prescription is a modular routing layer where models plug in and out, so a new release is a same-week test rather than a migration.
How do you deploy AI agents on a codebase without losing control?
Automated quality control on every commit, before any human looks at it. Mike's pipeline: agent commits, an automated checker reviews and can reject back to the agent, a human reviews only what survives, then commit, deploy, and DNS configuration execute automatically. Pair it with the guardrail principle that one agent should challenge another's output, because agents make confident mistakes the same way people do.
What should you never give an AI agent access to?
Secret keys and environment variables. Mike is unambiguous: keep the vault local. Working through public SSH keys is fine from a security standpoint. Handing credentials to a large external agent is not.
How do founders find their first customers for an AI product?
Mike's ranked answer: the healthiest path is an existing client base, because established trust means you can test freely and fix an early-stage product before it meets strangers. Second is partnership, pairing with someone who owns market access while you're the technical backbone. His active motion is Build With Not For, where an operator with clients and a proven method gets the method codified into a product with no upfront invoice, then a profit share if real users validate it.
Can AI actually replace a research team?
Partially, and the leverage is in parallelism. Mike reports running more than forty agents simultaneously to build a product landscape for a client's business plan, compressing months of human work into days. The structural insight holds regardless of the exact figure: the gain isn't one smart agent, it's many agents at once.
Should a founder run a services business and a product business at the same time?
Mike says yes, deliberately. Services alone means your only benchmark is client feedback, which is a sandbox rather than a market. Products alone means one win teaches you nothing about the next. Running both keeps each side honest, and the connective tissue is measurement.
How do you become a Claude partner?
You pitch, the same process as any other pitch. Joining the network starts free at the Registered level, with certification running through the Anthropic Partner Academy. Access brings learning materials, exams, and earlier visibility into upcoming releases.
URLs Mentioned in the Episode
- Saturnia Designhttps://saturniadesign.com
- Mike Vitez on LinkedInhttps://linkedin.com/in/mikevitez
- Ryan Estes on LinkedInhttps://www.linkedin.com/in/estesryan/
- AI for Foundershttps://aiforfounders.co
- Inbox Alchemyhttps://inboxalchemy.co
