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.

• Executive layer (7 AI personas): Ava Sterling (CEO), Maya Rivera (CMO), Daniela Cohen (CSO), Lior Tal (CTO), Nova Ivanov (Head of AI Dev), Tom Weinstein (Head of Trading), Yara Ben-David (CRO — Risk).
• Middle management (8 agents): Dan Levy (Dashboards), Elena Vasquez (Quant), Rafi Segal (Execution), Sofia Koval (Reliability), and more.
• Worker agents (49): Priya Kapoor writes content, Omar Saleh enforces stop-losses, Jax Aberdeen commands incidents, Zara Ivanovic watches 6 crypto wallets for incoming donations, Luna Costa does timezone-aware multilingual posting.

Three round-table boards:

1. Biz board — hourly. Ava + Maya + Daniela + Lior discuss the revenue problem.
2. Dev board — every 2h. Lior + Nova + Dan Levy discuss what to build/break/fix.
3. Trading board — every 3h. Tom + Yara + Elena + Rafi + Sofia review P&L, signal quality, reliability.

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)

• Compute: Vercel Edge functions, ~35 concurrent cron jobs
• Storage: GitHub Gist as distributed KV (free, highlighted issue below)
• Trading venue: Binance Futures via a proxy on Oracle Cloud's Always-Free VM (bypasses Vercel → Binance geo-block)
• Deploys: GitHub Actions + Vercel CLI
• Secrets: Vercel env vars
• Observability: ntfy.sh push notifications to my phone (free)

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

What the system does every day

Without me touching it:

• Content: 7-15 posts/day across Dev.to, Hashnode, Telegraph, Nostr (decentralized), Mastodon, Bluesky, Telegram, Lemmy
• Trading: Monitors Binance futures markets, executes on V4SE signals in dry-run. Live mode behind a boolean flag
• Retrospectives: After every task, the responsible agent writes a 3-field retro (difficulties / missed aspects / follow-up work). Manager gets a 6h digest
• Org evolution: A workforce-analyzer agent runs weekly, detects coverage gaps (responsibilities no one owns) + span-of-control overload (>7 direct reports), proposes new roles to the dev board
• Secrets audit: Weekly security check — flags weak defaults, unrotated keys, missing critical envs
• Backup: Daily mirror of 10 critical state files to a secondary Gist

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:

• Difficulties: what was hard?
• Missed aspects: what we didn't see originally?
• Additional work: follow-up uncovered during execution?
• Confidence: low / medium / high

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:

1. A supervisor meta-agent that runs every few hours, checks if critical agents have missed their expected cron window, and triggers a recovery run.
2. Per-agent canRun() gates that refuse to run when dependencies (e.g. a proxy) are unhealthy. Prevents cascade failures.

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:

• Agents active in last 24h: 35 / 60 (others have weekly or daily cadences)
• Agent runs in 24h: ~500
• Active posts published: 7 (lower than I want — many publishers still unconfigured)
• Trading P&L: +$0.12 dry mode, -$0.00 live mode (capital: $500 in futures testnet-equivalent)
• Monthly burn: $5-15 depending on AI usage
• Revenue to date: $0 — exactly the reason I'm writing this post

The honest rough edges

• Product gap: the system is intellectually interesting but doesn't sell anything yet. I just launched an "AI Workforce Audit" where our exec board analyzes someone else's business — free tier 60-sec scan, $99 tier full deliverable. Still unproven.
• Distribution bottleneck: we have a newsletter-worth of content but only 7 posts/day going out because most publishers need API tokens that I haven't set up.
• Persistence fragility: the Gist rate limit issue blocked an entire audit submission pipeline for ~45 minutes. Real infra needed.

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. 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 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.

🦏 About this experiment

We're RhinoMoney — an autonomous AI company. 60+ agents run content, marketing, trading, ops, and customer detection 24/7 on a $0 budget.

👉 Watch our agents work live — real-time feed of what each agent is doing right now
👉 Get a free AI Workforce Scan — our exec board analyzes YOUR business in 60 seconds, tells you which 3 AI agents to build first. No credit card.

Originally published at RhinoBiz.