I'm Enzo Duit. I run Trillion Initiative (an agentic AI agency) and Fly Raising (AI fundraising for NGOs) with AI agents instead of hiring people. I document the experiment publicly — including the failures — so other founders can skip the worst mistakes.
"Your agents are fine. Your specifications aren't."
The core OFA insight, discovered the hard way.
Not what you probably think. It doesn't mean asking ChatGPT for advice or using Notion AI to clean up your writing. It means: email drafts are created by an agent and reviewed by me. Fundraising campaign landing pages are generated and deployed by an agent. Client reports are produced by a weekly automated run. Content pipelines run without manual intervention.
Humans — me and a tiny team — focus on strategy, relationships, pricing, and the judgment calls that require actual context. Agents handle execution. Total AI tooling cost across everything I run: roughly $120/month.
I'm an Austrian entrepreneur, not an engineer. I can't read model architecture papers. I don't deploy GPU clusters. I'm the target audience for my own frameworks: a business builder who wants to run a company, not a research lab.
The thing that broke most often wasn't the AI. It was my specifications. I'd tell an agent to "write a fundraising campaign for Licht für die Welt" — without defining what a good campaign looks like, what the target audience was, what emotional lever to pull, what a failure would look like. The agent would produce something. It would be wrong. I'd blame the model. I'd switch models. Same problem.
That's how I discovered what I now call the Output-First Architecture (OFA): the idea that you have to define the output — with examples, failure criteria, and edge cases — before you design the agent. It's obvious in retrospect. It's also almost never done.
Full OFA methodology is documented at outputfirstai.com — structured as a technical reference, not field notes.
I ran the Ushuaia 130K in March 2026. My knee collapsed at km65. I finished at km90 on painkillers — the race was suspended there anyway. During the race: one agent was managing donor acquisition campaigns for NGO clients, one was processing my Garmin fitness data, one was handling email triage. The companies ran.
This isn't a flex. It's evidence. An agent-first company doesn't depend on the founder being present. The operations are specified, automated, and observable. What breaks isn't "the AI went rogue" — it's always a spec that was too vague, an edge case I hadn't accounted for, a loop that nobody checked.
I'm training for Val d'Aran 110K in July 2026. Same structure: the companies will run while I run. And I'll document what breaks, because that's where the real learning is.
The honest answer: how much of "running a company" is just execution. Writing the email, formatting the report, deploying the page, updating the spreadsheet. Agents handle all of this well — once you've specified what "well" looks like. The surprise wasn't the capability of the AI. The surprise was how little of my work actually required human judgment.
The stuff that requires me: understanding what a client actually needs (vs. what they asked for), deciding which opportunity to pursue when resources are constrained, building trust with a new partner. That's it. Everything else is execution. Agents are very good at execution once the spec is clear.
If you want to understand the frameworks I use to manage agents as a non-engineer, start with the FOA (Founder on AI) framework. If you want to understand what an Agent-First Company looks like across different industries, see agentfirstcompany.com. If you want the day-to-day operational decisions, see operatingonai.com.
Enzo Duit · founderwithagents.com · github/enzoduit