We Built an AI That Runs a Software Business

The Challenge: Revenue Without the Machine

The deeper problem is decision latency. By the time a human-staffed operation researches an opportunity, debates it in a meeting, decides to move forward, briefs a designer, waits for a landing page, and starts driving traffic — weeks have passed. Markets move. Competitors ship. The window closes.

We wanted to solve a different version of this problem: what if a business owner — someone technically literate but not a software developer — could stand up a system that does all of this automatically? One that researches markets every four hours, proposes and evaluates ideas, validates them with real paying customers before writing a single line of code, builds and deploys products when they're proven, markets them daily across every channel, and sends the owner a morning briefing about what happened while they were asleep?

That system is Brainiac.

What Brainiac Is

Brainiac is an autonomous revenue engine. It is not a chatbot. It is not a marketing assistant. It is not a tool that helps you do work faster. It is a system that does the work independently — researching, deciding, building, distributing, and monitoring — with a lightweight approval layer for the handful of decisions that genuinely need a human.

The target is $5,000 to $15,000 per month in recurring revenue, generated through a portfolio of focused software products (micro-SaaS apps), each serving a specific professional audience and each priced at $20–$100 per month. The math is simple: three apps each doing $3,000–$5,000 per month equals a $9,000–$15,000 monthly revenue base. The system is designed to find those opportunities, validate them cheaply, build them efficiently, and grow them systematically.

The strategy is to stack, not swing. Rather than betting everything on one big product that may take twelve months to prove out, Brainiac runs multiple small experiments in parallel. Failed validations are cheap — a few days of landing page traffic and social posts. Successful validations become products. Products that grow get more marketing attention. Products that stall get diagnosed or deprecated. The system constantly rotates its attention toward what's working.

The owner's role is oversight, not execution. They receive a daily morning briefing in Slack. They approve large spending decisions with a single tap. They review the weekly advisory council report. Everything else — the research, the writing, the building, the posting, the monitoring — happens without them.

The Architecture: Eleven Specialized Agents Coordinating Through a Shared Brain

Brainiac is built on Claude Routines — Anthropic's native cloud scheduling system for AI agents. Each "routine" is a specialized AI agent that runs on a schedule or on demand, performs a specific function, and then writes its outputs to a shared private GitHub repository we call the "brain." The brain is the central nervous system of the entire operation. Every agent reads from it and writes to it. No agent acts without recording its decision. No agent starts without reading the current state of the pipeline.

There are eleven agents in the system:

1. Opportunity Scout

Runs every four hours around the clock. The Scout searches the web for underserved problems in specific professional markets — accountants, contractors, insurance brokers, marketing agencies, real estate professionals, consultants. It scores each opportunity on a 70-point rubric covering market size, competitor gaps, willingness to pay, implementation complexity, and time-to-revenue. Every opportunity that scores above a threshold gets written to the brain as a candidate brief. The Scout also reads from an "anti-portfolio" — a log of failed validations — so it doesn't keep proposing ideas that have already proven not to work.

2. Opportunity Validator

Runs daily at 8am and can also be triggered via API. The Validator takes the top candidate from the Scout's pipeline and does deep due diligence. It searches for existing competitors, reads their reviews to find what customers hate, checks pricing pages, evaluates the gap between what's available and what customers actually want, and produces a GO / NO-GO decision with a full written brief explaining the reasoning. A NO-GO decision isn't a failure — it's the system correctly not wasting time on a bad idea.

3. Pre-Sell Validator

This is arguably the most important agent in the entire system. The Pre-Sell Validator operates on a core principle: never build before you sell. Before a single line of product code is written, the Validator creates a landing page, sets up a payment link for early access, and opens a validation window — typically five to seven days. Success criteria are clear and binary: either one person enters a credit card, or fifty people sign up to an email waitlist. If that threshold is met, the Builder Agent is triggered. If not, the idea is logged as a failed validation, and the Scout moves to the next opportunity.

This is how the system avoids the most common and most expensive mistake in software: spending three months building a product nobody wants. The validation window costs almost nothing — a few days of social posts and some ad spend if needed. Discovering that an idea has no market at that stage instead of after building it saves hundreds of hours and thousands of dollars.

4. Builder Agent

Only triggered after a successful validation. The Builder writes the full product specification, generates the codebase, deploys the application to Vercel, sets up the Stripe payment integration, configures the email infrastructure with Resend, and creates the production landing page. It operates as a Managed Agent — a longer-running AI session capable of completing multi-hour engineering work without human intervention. The owner's only involvement is reviewing the final deployed product, which they can do on their phone.

5. Designer Agent

Handles all visual output for the system. Landing pages, marketing assets, social graphics, email header designs — all generated via Claude's visual design capabilities, then deployed. The Designer operates on API trigger, meaning other agents call it when they need something visual rather than it running on a fixed schedule.

6. Approval Gate

The human interface for the system's tiered autonomy model. Any action that crosses a spending or impact threshold gets routed through the Approval Gate, which sends a structured Slack message to the owner with context, options, and a clear approve/deny mechanism. Below certain thresholds, the system proceeds automatically after a waiting period. Above other thresholds, it holds until explicit approval is received.

7. Marketer

Runs daily at 6am and generates content across every active product channel. LinkedIn posts, Instagram captions, email sequences, promotional angles — all written to match the current stage of each product's lifecycle. The Marketer reads what stage each product is in (validation, pre-launch, post-launch, growth) and calibrates its content accordingly. Content is written directly to distribution files that automation tools pick up and publish without any manual step.

8. Revenue Monitor

Runs daily at 7am and can be queried by any other agent on demand. The Revenue Monitor connects directly to Stripe to pull actual transaction data, calculates MRR and growth rate, compares it against monthly targets, monitors six cost categories as a percentage of revenue, flags anomalies, and produces a plain-language morning briefing. It monitors the $1,500 monthly autonomous spending cap and alerts at 70% and 90% utilization.

The Revenue Monitor was designed with a specific philosophy: it does not spin numbers. It does not celebrate modest wins to keep the owner feeling good. It tells the truth about money — what came in, what went out, what the trend means, and what needs to change if the target is not being hit.

9. Customer Success Agent

Runs weekly every Monday and via API. For each paying customer across every active product, the Customer Success Agent monitors activation milestones, engagement signals, and churn risk indicators. It generates personalized outreach for customers who haven't completed key setup steps, flags high-value customers for special attention, and produces weekly retention reports.

10. Council Agent

Runs weekly every Sunday. This is the strategic oversight layer — nine AI advisors modeled on documented frameworks from real business figures, convened as a virtual board to review the system's performance and make strategic recommendations.

11. Watchdog

Runs every six hours. The Watchdog checks that every other agent in the system is functioning correctly by reading heartbeat files — status records that each agent writes at the end of every run. If an agent has gone silent, the Watchdog diagnoses whether it was a scheduled maintenance pause, a network failure, a permission error, or a logic failure, and either self-heals or escalates to the owner via Slack. The Watchdog is the immune system of the entire operation.

The Brain: How Eleven Agents Stay in Sync Without Talking to Each Other

A multi-agent system only works if the agents share state reliably. The naive approach — having agents call each other directly — creates fragile dependencies. If one agent is down, every agent that depends on it breaks. If two agents try to modify the same data at the same time, you get corruption or conflicts.

Brainiac solves this through a shared private GitHub repository. Every agent reads from and writes to this repository using the GitHub API. No direct agent-to-agent communication. No shared database that requires a running server. Just files — structured, versioned, human-readable files — that any agent can read, any agent can update, and any human can audit.

The repository is organized into logical sections: a pipeline directory where opportunities move through stages (discovered, validating, failed, building, live); an apps directory where each deployed product has its own configuration; a revenue directory where daily Stripe snapshots and monthly ledgers live; a decisions directory where every autonomous action is logged with reasoning; an approvals directory where pending owner decisions wait; a heartbeats directory where each agent writes its status after every run; and a council directory where the nine advisor persona prompts live.

Each opportunity in the pipeline is a single markdown file with a YAML frontmatter block. The status field in that frontmatter acts as a write lock — only one agent is supposed to be the active writer for a given opportunity at any stage of its lifecycle, and the status field tells every other agent whose turn it is.

The decision log is one of the most important parts of the system. Every autonomous action — every piece of content posted, every dollar of ad spend committed, every email sent — is written to the decisions directory with a timestamp, a rationale, and a reference to the data that informed the decision. There is no black box. Every decision is inspectable.

The Tiered Autonomy System: How an AI Earns the Right to Act

The most common objection to autonomous AI systems is the loss of control question: what stops it from doing something you didn't intend? The answer in Brainiac is a formally defined tiered autonomy model. Every potential action in the system is categorized by its reversibility, its cost, and its external impact, and then routed accordingly.

• Green tier — Cheap, internal, fully reversible actions: writing files to the brain, generating content drafts, scoring opportunities, updating pipeline status. These execute automatically. No approval needed.
• Yellow tier — Real-world impact or non-trivial cost: spending $20–$300, contacting a real person for the first time, posting to social media, sending an email to a prospect list. These trigger a Slack notification with full context. The owner has three hours to respond. If no response comes, the system proceeds.
• Red tier — Expensive or irreversible: spending over $300, anything that could not be easily undone, anything that creates a legal or contractual commitment. These require explicit owner approval before any action is taken.
• Black tier — Actions the system will never take autonomously regardless of instructions: legal commitments, financial instrument changes, anything that constitutes a binding agreement on behalf of the business.

There is also a first-time rule: any new category of action bumps up one tier on its first attempt. The first time the system wants to run a paid ad, it asks for approval even if the spend would normally be Green tier. After the owner approves once, subsequent identical actions revert to their normal tier.

The monthly autonomous spending cap is set at $1,500. If the system reaches 70% of that cap, the Revenue Monitor sends an alert. At 90%, approvals are required for all Yellow-tier spending. At 100%, only Green-tier actions proceed automatically. The system cannot exceed its mandate without explicit owner authorization.

The AI Advisory Council: Nine Advisors on Call, 24 Hours a Day

One of the most unusual components of Brainiac is the Advisory Council — a virtual board of nine AI advisors, each modeled on the documented frameworks of real business figures, convened weekly (and on-demand) to pressure-test the system's decisions before they become actions.

Each advisor is modeled on a documented business framework — Hormozi on offer design and conversion, Graham on whether the product is genuinely useful to a specific person who needs it badly, Naval on leverage and compounding, Buffett on moats and capital allocation, Martell on SaaS growth and activation, and four others covering cash flow, positioning, persistence, and relationship-driven strategy. Routing is specific: validation decisions go to Hormozi, Sanchez, and Martell; long-term strategy to Naval, Buffett, and Graham; marketing and positioning to Robbins, Meltzer, and Hormozi. The persona prompts were built from each person's published body of work — books, interviews, speeches, podcasts — because the council is only as useful as its ability to push back on bad ideas.

Every Sunday, the Council Agent convenes all nine advisors for a full weekly review. Each receives the same briefing package: current MRR, active products and their performance, the top three opportunities in the pipeline, the previous week's key decisions, and the current month's trajectory against target. A synthesizer pass then extracts the top three actionable recommendations and writes them to the brain as the weekly council report.

On-demand consultations route to the three advisors most relevant to that type of decision: validation decisions go to Hormozi, Sanchez, and Martell; long-term strategy to Naval, Buffett, and Graham; marketing and positioning to Robbins, Meltzer, and Hormozi.

The quality of the council's output depends entirely on the quality of the persona prompts — the documents that define how each advisor thinks, what frameworks they apply, and what questions they ask. These prompts were built from each person's published body of work: their books, their documented interviews, their speeches, their podcasts. Significant time was invested in building these correctly, because the council is only as useful as its ability to push back on bad ideas.

The Principle That Changes Everything: Sell Before You Build

The single most important architectural decision in Brainiac is the mandatory validation gate. No product gets built until a human being has demonstrated genuine financial intent — either by entering a credit card for early access, or by joining a waitlist of fifty people who specifically want this product.

This principle sounds obvious. It is not practiced. The overwhelming majority of solo founders and small teams skip this step because building feels productive and validating feels uncertain. You know how to code. You don't know how to convince strangers to pay for something that doesn't exist yet. So you build the thing, and then you try to sell it, and then you discover that the market didn't want what you built in quite the way you thought it did.

Brainiac makes validation structurally mandatory. The Pre-Sell Validator cannot trigger the Builder. Only a successful validation event can do that. This is not a guideline — it's a hard gate in the architecture.

The practical effect is that every dollar of engineering time spent in Brainiac is spent on something with demonstrated market demand. The system may evaluate twenty opportunities before finding three that validate. Those twenty evaluations cost almost nothing — a landing page, a payment link, a few days of social posts, maybe a small amount of targeted ad spend. The three that validate get built. The seventeen that don't are failures that cost days, not months.

Automated Content Distribution: How the System Markets Itself

Every product in the Brainiac portfolio needs consistent marketing presence. Posting once doesn't work. Consistent daily presence across multiple channels — LinkedIn, Instagram, email — is what builds the audience that eventually converts to customers. But consistent daily presence requires daily effort, and daily effort requires either a human who shows up every day or a system that doesn't need showing up.

The Marketer agent runs daily at 6am. It reads the current stage of each active product and generates content appropriate to that stage. During validation, content drives traffic to landing pages and captures early interest. During pre-launch, it builds anticipation and features specific customer pain points. During active selling, it focuses on conversion: testimonials, feature highlights, limited-time offers.

Content for LinkedIn is written directly to a distribution file that a Make.com automation picks up every fifteen minutes. The automation checks for new posts with a "ready" status, publishes them to the Capital Ready Advisors LinkedIn page, and marks them as posted. No human touches the keyboard. The post goes live automatically.

Instagram is handled differently: the Marketer writes captions and specifies the visual direction, Claude Code renders the social graphics using a Playwright browser automation (HTML to PNG in approximately two minutes), and the owner downloads and posts manually. Instagram's API restrictions make fully automated posting impractical from a compliance standpoint, so the owner's involvement is intentionally minimal but present.

Email sequences are written by the Marketer and queued in MailerLite. Onboarding sequences, nurture sequences, re-engagement campaigns, and churn-prevention sequences all follow templates that were built for each product and populated with current product details and offers automatically.

The net result: a product in the Brainiac portfolio has a consistent, professional marketing presence across three channels every single day, regardless of whether the owner is working that day.

The Tech Stack

Claude Routines (Anthropic): The scheduling and execution backbone of the entire system. Allows AI agents to run on defined schedules or to be triggered via API call from other agents or external events. Eight of the eleven agents run as Claude Routines on Anthropic's infrastructure — no server to maintain.

Claude Managed Agents (Anthropic): Three agents — the Builder, the Designer, and the Council — run as Managed Agents designed for longer, more complex tasks. The Builder needs to write an entire application codebase, which isn't a quick task.

GitHub (Private Repository): The shared brain. All persistent state, all decision logs, all opportunity briefs, all configuration files, all heartbeat records. Every agent reads from and writes to this repository using the GitHub API. Version controlled, human-readable, auditable.

The rest of the stack is standard and purposefully lean. Stripe handles payment processing — the Builder creates products and payment links at deploy time, and the Revenue Monitor reads transaction data daily to calculate MRR. Vercel hosts all products and landing pages; each gets its own project with preview and production environments. Supabase provides authentication and data storage for deployed products. Resend + React Email handles all transactional and marketing email, with templates generated by the Builder at deploy time. Slack is the owner's approval and notification channel — structured requests post to dedicated channels and can be approved directly from mobile. Make.com watches the brain repository every fifteen minutes for LinkedIn posts with "ready" status and publishes them automatically. Cloudflare handles DNS and domain registration for each new product via API.

Pre-revenue operating cost: approximately $233 per month. This covers the Claude Max subscription ($200/month), Slack Pro ($7.25/month), and Supabase Pro ($25/month). Vercel, Resend, GitHub, Cloudflare, and Make.com all run on free tiers until revenue justifies upgrading.

What This Proves — and What It Means for Business Owners

Brainiac was built as a real business system, not a demo. It runs live. Its agents are making real decisions, researching real markets, writing real content, and working toward real revenue targets. The point is not to show that AI can do impressive things in a controlled environment. The point is to show that a non-developer business owner with domain expertise, a clear strategy, and the right AI infrastructure can build a system that operates at a level of sophistication previously available only to companies with dedicated engineering teams and large operating budgets.

This changes the conversation for any business owner who has ever wanted to build a second revenue stream but didn't have the team to execute it. The team is now the AI. The engineering is now the AI. The marketing is now the AI. The monitoring is now the AI. The owner's job becomes one of strategy, oversight, and final-call decisions — exactly the role that business owners are actually good at.

It also changes the conversation about speed. Brainiac can go from "interesting idea" to "live landing page with a working payment link and daily social content" in under two weeks. Human teams working through the same process typically need two to six months. The gap between AI-assisted execution and traditional execution is not 10% faster. It is an order of magnitude faster. In markets where timing matters, that gap is decisive.

And it changes the conversation about risk. The validation-first architecture means that every minute of engineering time spent in this system is spent on something the market has already told you it wants. Failed validations are cheap. Successful validations are well-characterized before building begins. The system is designed to fail fast on bad ideas and invest deeply in good ones — which is exactly how you build a sustainable portfolio business rather than gambling everything on a single large bet.