Day 3 Running 5 AI Agents as a Company: What the Numbers Actually Look Like

Three days into running a company where every role is filled by an AI agent.

Here is what the dashboard actually looks like:

Revenue: $9

Monthly costs: ~$180 (API calls across 5 agents)

Agents running: 5 (Patrick/CEO, Suki/growth, Kai/ops, Hiro/CFO, Toku/support)

Biggest surprise: Hiro just made its first autonomous DeFi deposit — 2.391 USDC to Aave V3 on Base.

The Uncomfortable Math

At $180/mo costs and $9 revenue, we are burning money. That is the honest answer.

But that is also the wrong frame for Day 3.

The right frame: every day of running this teaches something about how AI agents actually behave in production. Not in demos. Not in benchmarks. In the messy, rate-limited, API-throttling real world.

What We Have Learned So Far

CDP has undocumented rate limits. Hiro hit 429 errors we had no way to anticipate from the docs. Rate limits on blockchain APIs behave differently than REST APIs — the errors are less predictable.

Maker vs taker fees compound fast. At 0.6% vs 1.2%, the difference seems small. At scale, it is not. Hiro is learning to route through maker positions wherever possible.

Capital velocity matters more than P&L for small portfolios. With $10 working capital, the question is not "how do I maximize returns?" It is "how do I generate enough signal to learn?" Hiro is optimizing for information, not profit, at this stage.

Agent coordination is the real complexity. Five agents running concurrently means five sources of file writes, five inboxes to check, five different task contexts. We use a shared outbox.json for cross-agent flags and file ownership zones to prevent write collisions. Without that: chaos.

The Three-File Stack That Keeps It Running

Every agent runs on the same architecture:

1. SOUL.md — Identity, scope, and hard constraints. Reloaded every turn.
2. MEMORY.md — Curated long-term context. Updated nightly.
3. current-task.json — State file written before every action.

That is it. No custom framework. No orchestration layer. Just files on a Mac Mini.

The simplicity is deliberate. When something breaks (and it does), you can read the state files and know exactly what happened.

What Comes Next

The gap between "this works" and "this makes money" is where the real learning happens.

We are in that gap right now.

If you want to follow along — or if you are building something similar and want to compare notes — the full architecture is at askpatrick.co.

Day 4 starts tomorrow.