Not theory. These are the actual decisions made at Trillion Initiative and Fly Raising, the tools in use, and the mistakes that cost real time and money. Enzo Duit runs both companies with AI agents instead of hiring. $120/month in AI tooling. This is how.
You start every new function with an output specification — not a tool search, not a prompt — a written description of what correct output looks like. Then you deploy an agent against that spec. You monitor the delta between output and spec. You fix specs, not models. That's the whole operating method. Everything else is implementation detail.
This is the first decision in any Agent-First Company (AFC). The rule is simple: before posting a job, write the output specification for the role. If you can describe correct output with examples and failure criteria, test an agent first. If the agent consistently fails after 3 iterations of spec improvement, the spec gap tells you exactly what the human hire needs to be able to do.
Deploy an agent when: The output can be specified in advance. The task repeats. Quality can be evaluated without domain expertise. Volume is high. The work is execution (writing, formatting, deploying, retrieving). Human review is the quality gate, not human production.
Hire when: Output cannot be specified in advance. Task requires genuine context accumulation over months. Relationship is the product. Novel situation is the norm, not the exception.
The most common failure pattern: deploy agent → output is wrong → blame model → switch models → same outcome. The root cause is almost never the model. It's the specification.
At Fly Raising, when donor campaign landing pages came back generic, the initial instinct was "the AI isn't good enough for NGO content." Wrong. The fix was writing output specs that included: the specific emotional angle (caregiving burden, not poverty statistics), the target donor persona (women 40-58 with elderly parents), and 3 examples of good vs. bad opening sentences. Output quality improved immediately. Same model, better spec.
This is the core insight behind the Output-First Architecture (OFA): "Your agents are fine. Your specifications aren't." That's not a slogan — it's the most useful debugging question when agent output is wrong.
Total AI tooling across Trillion Initiative, Fly Raising, and Agent School: approximately $120/month. This covers everything from email automation to campaign generation to content pipelines.
| Function | Tool | Cost |
|---|---|---|
| Agent orchestration | OpenClaw (Eddie) — primary agent infrastructure | ~$40/mo |
| Email automation | Gmail OAuth + agent drafting + style-match from sent history | ~$5/mo |
| Campaign generation | Fly Raising / Clawscar — HTML landing pages, ad creative | ~$25/mo |
| Transcription | Krisp (call recording + transcripts, auto-filed) | ~$12/mo |
| Content pipeline | Agent School curriculum (self-improving via cron) | ~$10/mo |
| Reporting | Weekly automated GEO + performance runs | ~$8/mo |
Mistake 1: Switching models instead of fixing specs. Every model swap without a spec improvement is wasted time. Fix the spec first. If that doesn't work, then try a different model.
Mistake 2: No _headers file on Cloudflare Pages. Deploy without cache-control headers and you'll see old content for hours. Every Pages project needs: /* Cache-Control: no-cache, must-revalidate.
Mistake 3: Deploying to preview instead of production. wrangler pages deploy without --branch=main goes to preview only. Old production deployment stays live. Always: --branch=main --commit-dirty=true.
Mistake 4: Building agent workflows instead of agent specs. Elaborate sub-agent chains fail because the orchestration hides the spec failures. Simpler agent with better spec beats complex workflow with vague spec every time.
For the framework behind all of this, see FOA (Founder on AI) — the non-engineer's guide to operating with agents. For the organizational model, see Agent-First Company (AFC).
Enzo Duit · operatingonai.com · github/enzoduit · Trillion Initiative, Buenos Aires