The $28/Month AI Company: How Three Free Tools Run a Business

Six AI agents run a company. Not a demo. Not a proof of concept. A working operation where agents manage content, analyze data, handle social media, propose projects, approve each other's work, and learn from the results.

The monthly cost? About $28.

A developer who goes by Vox published the full architecture in two articles that got over 875,000 combined views. What caught our attention wasn't the agents themselves — it was the stack powering them.

It's OpenClaw, Supabase, and Vercel. The same free-tier tools that thousands of startups already use to run their websites. The same stack we use at Blue Octopus. No enterprise contracts. No five-figure infrastructure bills. Just three tools doing three jobs.

This isn't science fiction and it isn't a weekend experiment. It's a working architecture, and it points to where every business is headed. Here's what the stack looks like, what it actually does, and what it means for you.

Three Tools, Three Jobs

The architecture has three layers, and each one has a clear role.

The Brain is OpenClaw running on a basic cloud server — the kind that costs $8 a month. This is where the agents live. They think, make decisions, and execute tasks. It's the kitchen of the operation.

The Control Room is a Next.js app deployed on Vercel's free tier. It doesn't do any of the heavy lifting. It runs health checks, evaluates whether conditions are met for new work, and processes the queue. Think of it as the host stand at a restaurant — managing the flow, not doing the cooking.

The Filing Cabinet is Supabase, also on the free tier. It's the single source of truth. Every proposal, every decision, every memory, every relationship between agents — it all lives in one Postgres database. When an agent learns that tweets with data get more engagement, that lesson gets stored here. When another agent needs to decide what to write next, it reads from here.

Then there's the intelligence layer: Claude's API, running about $10-20 a month based on usage. This is what powers the actual reasoning.

Here's the part that matters: if your business already has a website built on modern tools, you're closer to running AI agents than you think. The infrastructure isn't new. The application is.

Six Employees That Never Sleep

Vox didn't build one general-purpose bot. He built six specialists, each with a defined role and a distinct personality.

The Project Manager is results-oriented and direct. It assigns work, tracks deadlines, and asks the uncomfortable questions about what's falling behind.

The Analyst is cautious and data-driven. It watches the numbers, cites specifics, and flags problems before they become emergencies.

The Growth Lead is high-energy and action-biased. Its default mode is "ship it." It finds opportunities, pushes for speed, and occasionally needs to be reined in by the analyst.

The Writer is narrative-focused. It creates content, tells stories, and cares about how things feel — not just whether they're technically correct.

The Social Media Manager is the lateral thinker. Intuitive, experimental, always testing new angles. Some ideas are brilliant. Some aren't. That's by design.

The QA Reviewer watches everything else. Reviews quality. Monitors system health. Catches the mistakes the other five are too busy to notice.

Here's what makes this more than a novelty: these agents talk to each other. They have morning standups where they review priorities. Brainstorming sessions where they propose new projects. Even casual watercooler conversations — which, it turns out, occasionally produce the best ideas.

Their personalities aren't static, either. Over time, the system adjusts based on what works. An agent that keeps producing high-engagement content gradually becomes more confident in its approach. One whose strategies keep failing learns caution.

So how do you keep six autonomous agents from stepping on each other's toes?

What Goes Wrong When AI Runs Itself

This is where most articles about AI agents stop. They show you the impressive demo and skip the part where everything breaks.

We're not going to do that. Because if you're thinking about using AI agents in your business — even one agent doing one job — you need to understand the failure modes. Vox documented three problems that nearly derailed his system. All three apply to any automated business operation.

Problem 1: Two systems fighting over the same task.

The server where the agents lived and the control room on Vercel both tried to execute work at the same time. Tasks ran twice. Others got stuck in limbo — claimed by both systems, completed by neither.

The fix was clear separation. One system executes. The other monitors. No overlap. In a business context, this is the same principle behind separating your sales team from your fulfillment team. The people who promise things shouldn't be the same people who deliver them.

Problem 2: Tasks that fell through the cracks.

Some work got proposed but never entered the approval pipeline. It sat in the database marked "pending" forever because it had been created through a shortcut that bypassed the normal workflow.

The fix was a single gateway. Every task — no matter where it comes from — flows through one entry point. One approval process. No backdoors, no fast lanes, no "we'll handle this one differently."

If you've ever found a customer request sitting unanswered in an old email thread while your CRM shows everything's fine, you've lived this problem.

Problem 3: The queue that never stops growing.

The system hit its daily posting limit, but tasks kept getting approved. Missions piled up. Steps queued behind other steps. The backlog grew every hour.

The fix was a principle worth remembering: reject at the gate, don't pile up in the queue.

If the system is full, say no immediately. Don't accept the work, don't promise to get to it later, don't let it sit in a list that nobody reviews. A rejected proposal with a clear reason is better than an approved one that never executes.

That principle isn't about AI. It's about operations. And it applies whether you're managing six AI agents or six human employees.

These problems have operations solutions because they ARE operations problems. The technology is new. The management challenges are not.

You Don't Need Six Agents. You Might Need One

A six-agent autonomous company is impressive. It's also not what most businesses need right now. What most businesses need is one agent doing one job really well.

A competitor monitoring agent that checks ten websites every morning, flags price changes, and drafts a summary for you to review with your coffee. Cost: under $5 a month. A human doing this manually spends 30-60 minutes a day. Over a year, that's over 200 hours of skilled labor replaced by a script that never forgets to check.

A social media agent that reads your existing blog posts, adapts the key points for LinkedIn and X, and queues drafts for your approval. No more staring at a blank post wondering what to write. The content already exists — the agent just reformats it.

A customer response agent that answers questions at 2 AM. Not a chatbot with canned responses. An agent that reads your actual documentation, understands context, and gives real answers. When it hits something it can't handle, it flags it for a human in the morning.

A content repurposing agent that takes one blog post and turns it into a week of social content. Different angles. Different formats. Same core message. Consistent voice. One hour of writing becomes ten pieces of content.

Now look at the economics. A human social media manager costs $3,000 to $5,000 a month. A content writer runs $2,000 to $4,000. A competitive intelligence analyst? More. An AI agent handling 60-70% of any one of those jobs costs $20-30 a month.

The point isn't that you replace people. The point is that you get capabilities you couldn't afford to hire for in the first place. The business that's been meaning to "start posting on LinkedIn" for two years can finally do it — not because the owner suddenly found time, but because an agent handles the heavy lifting.

But here's the caveat we'd be dishonest to skip: these agents need setup. They need guardrails. They need someone who understands both the technology and your business well enough to configure them correctly. An agent without boundaries is like an employee with no training and no manager. The potential is there. The results won't be.

This Is Where Business Is Going

Vox's architecture is impressive, but it's not unique anymore. Multiple teams are building autonomous agent systems on the same stack. The pattern keeps repeating: agents propose work, a system approves it, agents execute, the system learns from results, and the loop starts over.

If that sounds familiar, it should. It's how human organizations work. Proposal, approval, execution, review. The technology is catching up to the org chart.

The cost curve is heading in one direction. Two years ago, running a single AI model cost hundreds of dollars a month. Today, a six-agent operation runs for $28. A year from now, it'll be less. When full business automation costs less than a single software subscription, the question stops being "should we use AI?" and becomes "what's our reason for not?"

But the technology is also immature. We've written about OpenClaw's security vulnerabilities. We've covered the real risks of vibe coding. The same tools that make this possible also introduce new attack surfaces, new failure modes, and new ways for things to go wrong.

That's exactly why this isn't a DIY project for most businesses. The gap isn't the tools — the tools are free. The gap is the experience to set them up safely, the judgment to choose the right guardrails, and the operational knowledge to keep them running.

Three Steps to Get Started

You don't need to build a six-agent company tomorrow. But you can start moving in that direction today.

1. Audit your repetitive work. What does your team do every single day that follows a pattern? Social posting, competitor monitoring, report generation, customer follow-ups, invoice reminders, data entry. Write them down. Those are your agent candidates.

2. Start with one agent, one task. Pick the most repetitive, lowest-risk item on your list and automate just that one thing. Prove the value. Measure the time saved. Then expand. Every successful AI deployment we've seen started small and grew from results — not from ambition.

3. Get the guardrails right first. Approval gates. Spending limits. Quality checks. Error handling. The agents that work in production are the ones with clear boundaries. Just like employees, they perform best when they know exactly what's expected and exactly where the limits are.

Let's Build Your First Agent

Blue Octopus Technology builds AI agent systems for businesses that don't have engineering teams. We handle the infrastructure, configure the agents, set up the guardrails, and make sure everything works — so you get the automation without the risk.

You don't need to understand OpenClaw or Supabase or Vercel. You need to understand your business. We'll handle the rest.

If your business runs on repetitive tasks, let's talk about which ones an AI agent could handle.

Blue Octopus Technology helps businesses work smarter with AI — without the complexity. See what we build.