🦄 One Person Unicorn
Submit Your Company →Submit

playbook · Julien de Waal · 11/19/2026 · 5 min read

Building with an Agent Fleet: What Actually Works (and Breaks) for Solo Founders

Mike Rice didn't hire a team. He built a fleet.

The San Francisco-based founder of AGI Future Foundation maintains a database of 303 verified startup programmes — accelerators, grants, fellowships, pre-seed funds — using a small group of AI agents. That database powers a free one-apply tool built specifically for bootstrapped founders who don't have time to research 300 application portals.

One person. One product. Operational infrastructure that runs largely without him.

That's not a side project. That's a blueprint.

What the agent fleet actually does

Rice's agents handle what most early-stage teams split across two or three hires: data sourcing, verification, maintenance, and quality control. The 303-programme database isn't static. Deadlines change. Programmes close. New ones open. Keeping it accurate is exactly the kind of repetitive, high-volume work that burns out junior employees and bores senior ones.

Agents don't get bored.

The architecture isn't described in granular detail in his post, but the structure implies a few distinct agent roles: one or more scrapers or monitors tracking source data, a verification layer checking accuracy, and presumably some coordination logic deciding when human review is actually needed. That's a standard multi-agent workflow pattern — specialised agents with narrow jobs, passing outputs between them, with a human in the loop only at exception points.

This is the operational model that defines the emerging class of AI-native companies: not companies that use AI as a feature, but companies where AI agents are the operational layer itself.

Where it breaks

Rice is upfront about failures, which is what makes his account worth reading. Agent-run infrastructure doesn't break dramatically — it breaks quietly. A scraper hits a changed page structure and returns nothing. A verification agent flags something incorrectly. The database drifts from reality in ways that aren't immediately visible.

This is the core operational risk of agentic systems: silent failures. A human employee doing this job would notice when something felt off. Agents don't have that instinct. They execute the task they were designed for, and if the task definition no longer matches reality, they execute the wrong thing reliably and at scale.

The practical implication: monitoring is not optional. Any agent fleet doing real operational work needs exception handling, alerting, and regular human audits — not to micromanage the agents, but to catch the moments when the environment has changed and the agents haven't adapted.

Rice's experience also points to a less-discussed challenge: context windows and stateless execution. Agents working across a large dataset don't naturally maintain awareness of what they've already processed, what's changed, and what dependencies exist between records. Solving that requires deliberate architecture decisions — shared memory layers, structured handoffs, or database states that agents can read before acting.

The economics behind the model

There's a reason the revenue per employee metric has become the defining number for this generation of AI-native startups. Rice's model is a clean illustration of why.

Maintaining a 300-entry verified database with regular updates, cross-referencing sources, and a production-ready search interface would historically require at least one full-time operations hire, probably two if the product is meant to stay accurate. At San Francisco market rates, that's $150K–$200K in annual salary before benefits. With agents, that cost compresses to API calls, infrastructure, and the founder's time spent on oversight and prompt engineering.

The product is free — Rice isn't monetising the database directly. But the model demonstrates what happens when operational labour costs approach zero: products that couldn't be profitable at traditional headcount costs become viable. That's not a marginal efficiency gain. It's a structural shift in what a solo founder can build and sustain.

What solo founders can take from this

Rice's notes are operational, not theoretical. A few patterns worth extracting:

Narrow the agent's job. The most reliable agents in a fleet are the ones with the smallest surface area. One agent monitors one type of source. One agent validates one field type. When something breaks, you know exactly where to look.

Design for drift. Databases go stale. External sources change. Build your agents assuming the data environment will shift, and build your monitoring assuming agents won't notice when it does.

Human review is a feature, not a failure. The goal isn't a fully autonomous system — it's a system where your time is spent on decisions only you can make, not on tasks agents can handle. Keeping humans at exception points, rather than eliminating them entirely, is what makes the fleet actually trustworthy.

Start with the most repetitive work. The 303-programme database is maintained because the maintenance work is predictable and high-volume. That's the right starting point for any agent fleet. Don't start with judgment-heavy work. Start with the work that would otherwise consume hours of someone's week doing roughly the same thing over and over.

This is the same logic behind what Julien de Waal demonstrated at SwissBorg: 300 SEO pages produced in a single quarter, app installs scaling from 600 to 25,000 in three months — results driven by agentic systems handling the volume work that content and growth teams would otherwise staff for.

The solo founder infrastructure stack is maturing

A year ago, building reliable agent-run infrastructure was largely an exercise in prompt engineering and debugging. The tooling is catching up. Frameworks like LangGraph, CrewAI, and AutoGen have made multi-agent architectures more accessible. Hosted memory layers, structured output enforcement, and agent observability tools mean that the failure modes Rice describes are increasingly solvable with off-the-shelf components rather than custom code.

The founders who are figuring this out now — building real products on real agent infrastructure, documenting what breaks, and iterating — are accumulating operational knowledge that won't be easy to replicate when the tools become mainstream.

Rice's one-apply tool for bootstrapped founders is a small product. But the infrastructure model behind it is the same one underpinning the one-person unicorn thesis: that a single founder with the right agent architecture can sustain operational complexity that previously required a team.

The fleet doesn't replace the founder. It replaces the headcount the founder never hired.

---

Is your company eligible? Submit to the leaderboard → onepersonunicorn.co/submit

Read the full AI-native companies guide.

Is your company eligible? Submit to the leaderboard →

Submit Your Company

More on AI Agents for Founders: The Complete 2026 Guide

How a Solo Founder Replaced a Document Processing Team With a Multi-Agent AI PipelineSoGood.ai Review: Can One Platform Replace Your Entire Founding Team?One Developer vs. the Bot Internet: How Solo Founders Are Solving Human Verification Without Big Tech

Related companies on the leaderboard

Sonscape

Undisclosed ARR · —

Polsia

$1M ARR · $1M/person

Swan

$1M ARR · $333k/person