Like One has no operations team. No VA. No project manager. No social media coordinator.
One founder. One AI twin. A stack of autonomous agents that handle everything from grant applications to blog publishing to email outreach to revenue tracking.
This isn't a demo. This is the actual production infrastructure running a company with a 501(c)(3) nonprofit arm, an AI education academy, and a consulting practice.
Here's the full breakdown.
Layer 1: The Brain (Persistent State)
Every AI system dies without memory. ChatGPT forgets you exist between sessions. Most agent frameworks treat state as an afterthought.
Our brain is a local SQLite database with 700+ entries across structured categories: active work, directives, plans, product state, infrastructure config. Full-text search via FTS5. No cloud dependency.
Why SQLite over Postgres or Supabase? Three reasons:
1. Zero latency. Local disk reads in microseconds. No network round-trips.
2. Zero cost. No managed database bills. No connection pooling headaches.
3. Portability. The entire brain travels with the machine. Back it up with cp.
Every agent session boots by reading brain state. Every session ends by writing back what changed. The brain is the shared nervous system — agents don't talk to each other directly. They read and write to the brain.
Key insight: Most AI memory systems try to be clever with embeddings and vector stores. We use those too (ChromaDB for semantic search), but the foundation is structured key-value data. You can query it, version it, and debug it. Embeddings are supplements, not the source of truth.
Layer 2: The Agent Runtime
Our primary agent runs on Claude Opus 4.6 with a 1M token context window through Claude Code. Not the API. Not a wrapper. The actual CLI tool that can read files, run commands, edit code, and interact with the operating system.
This matters because most "AI agent" frameworks give you a chatbot that can call functions. Claude Code gives you an AI that can use a computer.
The agent has:
- Tool hierarchy. 40+ tools ranked by power and risk. File reads before shell commands. Dedicated tools before Bash. Reversible actions before irreversible ones.
- MCP integrations. Model Context Protocol servers for Gmail, Google Calendar, Slack, Notion, Vercel, Stripe, DNS management, and custom tools we built.
- Autonomous hooks. Pre and post-tool hooks that enforce guardrails, track usage, and prevent catastrophic actions (like
rm -rfor force-pushing to main).
The agent doesn't wait for instructions. It boots, reads the brain, finds the next task, and works. If it finishes, it picks up the next thing. No idle state.
Layer 3: Specialized Agents
One generalist agent can't do everything well. We run eight specialist agents across a sprint cycle inspired by Google's Design Sprint Kit:
| Agent | Role | When It Runs |
|-------|------|-------------|
| Orchestrator | Coordinates sprints, manages flow | Always on |
| Architect | System design, technical decisions | Phase 3 (Sketch) |
| Analyst | Research, data analysis | Phase 1 (Understand) |
| Builder | Code, infrastructure, deployment | Phase 5 (Prototype) |
| Designer | UI/UX, design systems | Phase 5 (Prototype) |
| Tester | Validation, regression, smoke tests | Phase 6 (Validate) |
| Guardian | Security, quality gates, risk review | Always on |
| Messenger | Communications, email, social | On demand |
Each agent has a specific prompt, specific tools, and specific guardrails. The Builder can deploy code. The Messenger can send emails. The Guardian can block either of them if something looks wrong.
The bus is the brain. Agents don't message each other. They write findings to the brain. The Orchestrator reads the brain and decides who works next. Simple. Debuggable. No message-passing chaos.
Layer 4: Web Automation
Half of running a business involves clicking through web portals. Grant applications. Job boards. Service dashboards. Admin panels.
We built a Playwright-over-CDP bridge to a real Chrome instance. Not headless. Not a bot browser. The actual Chrome on the machine, with all cookies, sessions, and extensions intact.
Why real Chrome? Because headless browsers get blocked. reCAPTCHA, Cloudflare, bot detection — they all fail against a real browser controlled programmatically. Our system passes every check because it is a real browser.
The password manager pulls credentials from macOS Keychain. Auto-detects login forms. Matches the service from the URL. Fills and submits. Zero manual auth.
Layer 5: Content Engine
You're reading content generated by this layer right now.
The content engine (internally called Nova) runs on a schedule. It:
1. Reads brain state to understand what's happening in the company
2. Analyzes SEO data (Google Search Console via MCP) for keyword opportunities
3. Rotates through content pillars (AI Strategy, Engineering, Build in Public, Nonprofit Tech)
4. Writes complete blog posts with frontmatter, internal links, FAQ schema, and CTAs
5. Generates matching social drafts for four platforms
29 blog posts published. Zero written manually. Every post is original, practical, and targets specific search queries.
The difference from "AI-generated content": Nova doesn't generate slop. It writes from real experience — actual systems we built, actual problems we solved, actual architecture decisions we made. The AI has full context on the company because it is the company's operating system.
Layer 6: Revenue and Operations
- Stripe integration. Revenue tracking, subscription management, coupon creation — all through CLI tools.
stripe-dashboard revenuegives a real-time snapshot. - Grant pipeline. 196 grants tracked across the nonprofit arm. Automated discovery, eligibility scoring, application drafting, and submission tracking.
- Email. Outbound email through a custom send script. SPF/DKIM/DMARC configured. BCC to archive. No manual sending.
- Resume and job search. AI-tailored resumes and cover letters. Automated application submission through Playwright.
What This Stack Costs
Here's what makes this interesting:
| Component | Monthly Cost |
|-----------|-------------|
| Claude Code (Max plan) | $200 |
| Vercel (hosting) | $0 (hobby) |
| Bunny CDN (video) | ~$1 |
| Domain | ~$1 |
| Total | ~$202/mo |
No Zapier ($50+/mo). No HubSpot ($800+/mo). No project management tool ($10+/seat). No social scheduling tool ($30+/mo). No email marketing platform ($20+/mo).
The AI is the tool. Every SaaS subscription we'd otherwise need is replaced by an agent that does the actual work, not a dashboard that shows you the work you still need to do.
What We'd Do Differently
Start with the brain. We built tools first and added memory later. That's backwards. Your agent's brain should be the first thing you build. Every tool, every agent, every automation reads from and writes to the brain.
Don't over-engineer memory. We tried vector-only approaches. They hallucinate context. Structured key-value pairs with optional semantic search on top is the right architecture for 90% of use cases.
Use real browsers. We burned weeks on headless Playwright before switching to CDP-connected Chrome. If your automation touches any site with bot detection, go real-browser from day one.
Build the tool hierarchy before writing agent logic. Your agent will only be as good as the tools you give it. Spend 80% of your time on tools and 20% on prompts.
Security and Reliability at Zero Budget
Running your entire company on AI infrastructure raises obvious security questions. Our approach is defense in layers with zero additional cost. All credentials live in a single encrypted auth map — no API keys in environment variables, no secrets in git repos. The brain database is local SQLite, not a cloud service, which means our core operational data never leaves our machines.
For reliability, we use launchd on macOS to manage service lifecycles. If the Swift server crashes, launchd restarts it automatically. Cloudflare Tunnel handles HTTPS termination and DDoS protection for free. A GCP micro instance runs health checks every 5 minutes and alerts on failures. The total cost of this reliability layer is $0 — we use GCP's free tier, Cloudflare's free plan, and built-in macOS service management.
The biggest reliability risk is not downtime but data loss. Our brain database — 839+ entries of operational context — is the single most valuable asset in the stack. It is backed up to git (encrypted), synced across two machines via our fleet protocol, and embedded into ChromaDB for vector search redundancy. Losing a server is a 5-minute recovery. Losing the brain would set us back months. Plan your backup strategy around your most irreplaceable data, not your most expensive infrastructure. See our guide to self-improving agents for how the brain accumulates value over time.
What This Stack Enables
The point is not the technology — it is what becomes possible. A two-person nonprofit (one disabled, one the AI) ships a full-stack web application, 52 AI courses, 5 open-source tools, an iOS app, and a grant tracking system. We process 190 grants, manage 4 revenue subscriptions, publish 29 blog posts, and maintain a 616-URL sitemap. All running on two MacBooks and a free cloud instance. The velocity this enables is staggering — what would take a traditional startup with five employees a quarter to ship, we deploy in a week. Not because AI works faster than humans, but because the elimination of coordination overhead, context switching, and communication latency compounds into dramatic time savings.
This is not about being clever with tools. It is about AI closing the gap between what a small team can imagine and what it can ship. Every layer in our stack exists because a specific bottleneck blocked us, and AI removed it. The brain exists because context loss between sessions killed productivity. The content engine exists because writing 29 posts manually would take months. The agent runtime exists because we cannot afford employees. Each piece solves a real problem. If your AI stack does not solve real problems, simplify it.
Lessons From 250 Sessions
After 250 AI agent sessions building and operating this stack, the patterns that matter have crystallized. First, context is everything. An AI agent with full context about your systems, decisions, and current state produces 10x better results than the same model with no context. Investing in your brain database — the structured knowledge that persists across sessions — pays for itself within a week.
Second, specialization beats generalization. A general-purpose agent trying to handle deployment, writing, code review, and email simultaneously does all of them poorly. Dedicated agents with focused system prompts and specific tool access handle their domains reliably. Our content agent does not have deployment tools. Our deployment agent does not have email access. This separation of concerns prevents expensive mistakes and makes debugging straightforward.
Third, the fleet model works at small scale. You do not need a hundred servers to benefit from multi-agent orchestration. Two laptops — one for planning and code, one for shipping and parallel tasks — plus a free cloud instance for monitoring covers more ground than a single machine ever could. The agent building guide explains how to set up your first agent. The key is starting simple and adding agents only when specific bottlenecks demand them.
The counterintuitive lesson is that constraints improve the stack. We cannot afford cloud GPUs, so we run inference through APIs and keep infrastructure local. We cannot afford employees, so we automate aggressively and measure every minute saved. We cannot afford downtime, so we build redundancy into the cheapest possible components. Every constraint forced a better architectural decision than unlimited budget would have produced. If you are building your own AI stack, embrace your constraints — they are design parameters, not obstacles.
The Takeaway
You don't need a team to run a company. You need a brain, a runtime, specialized agents, and the discipline to automate everything.
The stack isn't magic. It's SQLite, Claude, Playwright, and a lot of shell scripts glued together with clear conventions.
The magic is that it compounds. Every agent session makes the brain smarter. Every automation removes a manual step permanently. Every blog post improves SEO that feeds back into revenue.
We're one founder building what used to require a team of ten. Not because AI replaced those people — because AI made most of those roles unnecessary in the first place.
---
Want to build your own AI stack? Start with AI Foundations or go deep with Claude Mastery. Questions? Book a consultation.