60 AI agents are running a real business right now. I'm not touching it.

Six weeks ago I asked: can a fully autonomous AI workforce actually operate a real business end-to-end? Not a "copilot." Not a ChatGPT wrapper. Full autonomous operation — content, marketing, customer detection, trading, monitoring, even exec-level decision-making.

Today there are 60+ specialized AI agents running RhinoMoney. Three executive boards meet every few hours to deliberate priorities. Agents publish content across 8 platforms. Trades get placed. Retrospectives get written. An org chart manages itself.

This post is a receipt, not a pitch. Architecture, numbers, failures.

The org chart

Every agent is an entity with a name, a manager, KPIs, and a voice. Not anonymous functions.

Three round-table boards:

Each board ends with a CEO decision. That decision is AI-parsed into stories + action items. An `action-executor` agent picks up action items every 10 minutes and routes them to worker agents.

The stack (free tiers all the way down)

Total paid services: OpenAI + Claude API. Current burn: ~$5-15/month.

What the system does every day

Without me touching it:

The interesting architectural pieces

1. Boards write to a shared "Actions Board"

Each board's CEO decision is parsed into stories with action items. Stories are typed: `{backlog, planned, in_progress, review, done, blocked}`. Action items have assignees (agent name or human). Every agent run that completes an action writes a retro back onto it.

This means the boards cross-pollinate. When the biz board decides "we need more distribution," the dev board sees that open action at its next meeting and can respond with technical proposals.

2. Every agent is in an org chart

Before: agents were a flat dict. After a week I had 40+ agents with no clear ownership, no KPIs, no way to tell if coverage had gaps.

Now: `org-chart.ts` models every agent as a member with `reports_to`, `manages`, `responsibilities`, `kpis`. A weekly analyzer does `analyzeCoverage()` — matches responsibilities against a list of critical-but-often-missed concerns (security audit, backups, legal compliance, cost monitoring). If nothing in the org owns "GDPR/privacy" — it's a coverage gap. A new role proposal is auto-posted to the dev board.

Caught gaps this way: "legal compliance" (fixed with /terms /privacy /disclaimer + a compliance-checker agent), "backup/restore" (fixed with backup-agent), "customer support post-sale" (fixed with customer-success agent that polls Gumroad + Shopify).

3. Retrospectives happen to the AI, by the AI

After an action completes, a `status-synthesizer` agent prompts the original worker agent (in its persona) to write a retro:

This gets attached to the action item. Every 6h, retros roll up into a manager digest I get as a push notification.

Surprisingly useful. Actually reads like a team debrief.

What broke

Not a highlight reel.

GitHub Gist rate limit (5000/hr). Thirty-plus agents reading state every few minutes plus writes turned out to be enough to blow through this. Symptom: writes silently failed for 60 minutes until the hourly reset. Fix: extended in-memory cache TTL 20x (from 3s to 60s), slowed non-critical crons, now comfortably under the ceiling. Real fix next: migrate to Upstash Redis (10k/day free, Edge-native).

Facebook Groups API gone. Meta deprecated it in April 2020. No legitimate way to post to groups via API. Workaround: a Chrome extension that runs in my own signed-in Chrome session, opens tabs to target groups, fills composers with human-paced typing (40-160ms per character, 1-3s hovers, random delays), submits. Meta can't distinguish these from manual use. Throttled to 2 posts/day across 15 curated groups.

Agents that silently fail. Fixed twice:

Spending going unnoticed. Added a `burn-monitor` agent that estimates monthly spend from run counts and alerts at $5 / $15 / $30 / $50 / $100 thresholds.

Numbers I'm comfortable sharing

From the current week's run:

The honest rough edges

Why this matters (to me)

I think autonomous AI agents will eat whole categories of small business operations in the next 12-24 months. Not "AI copilot" — actually autonomous. This is my attempt to find out what that looks like when you push it to the limit.

If you run a business and want to see where AI agents could replace your grunt work, I built a free 60-second scan: [https://rhinomoney.app/audit/apply](https://rhinomoney.app/audit/apply). Our actual exec board analyzes your submission and returns 3 ranked opportunities with realistic ROI estimates.

Or just watch the system work: [https://rhinomoney.app/live](https://rhinomoney.app/live) shows every agent run in real-time.

Either way — this is an open experiment. If it's interesting, I'm posting the code + weekly write-ups on the blog. All free. Receipts included.