The AI Agency Blueprint

QUICK START

Build Your First Agent in 30–45 Minutes Before reading anything else, let’s build one agent and see it work. A real result in under an hour — then the rest of the guide makes sense because you’ve already done it once. First-time Gmail OAuth may add 15 minutes on top.

WHAT YOUʼ LL HAVE AT THE END

Maya — your Chief of Staff — running as a scheduled task, pulling your Gmail inbox, and producing a structured morning brief. One agent. Fully operational. Step 1 — Connect Gmail MCP (10 min)

01 Open Claude → Settings → Integrations

02 Search “Gmail” → Connect → sign in → approve permissions 03 Verify: in a new conversation, type “List my 5 most recent emails” — Claude should return real results Step 2 — Create Maya’s Scheduled Task (15 min) 01 Open the Cowork desktop app (download at claude.ai/download)

02 Click Scheduled Tasks → New Task

03 Name: maya-morning-brief 04 Cron: 0 9 1-5 (9 AM Mon–Fri) 05 Paste this starter prompt into the task:

MAYA STARTER PROMPT — COPY & MODI FY

You are Maya, Chief of Staff for [YOUR BUSINESS NAME]. Every morning, check Gmail for emails received in the last 18 hours. Summarize: (1) unread messages requiring action, (2) client or prospect replies, (3) anything flagged urgent. Format:

MORNING BRIEF — [DATE]

ACTION REQUIRED: [list] CLIENT/PROSPECT UPDATES: [list] URGENT FLAGS: [list] Use gmail_search_messages to search recent emails. Send the completed brief to [YOUR EMAIL] via Gmail. 01 Click Run Now on the task 02 Claude asks permission to use Gmail — click Approve 03 Watch the output — you should see a real morning brief appear 04 If it works: enable the task for automated daily runs ⚠ Cowork-based scheduling has a hard dependency: your laptop must be on, awake, and running Cowork for tasks to fire. This is the single most fragile assumption in this guide. If you close the lid, tasks stop. Travel for a week, and your agency goes dark. For revenue-critical workflows, see Ch. 4.5 (Worker orchestrator). “Done. You’ve built your first operational AI agent. Now read the rest of this guide to build the other seven.”

DO THI S NEXT

Read chapters 1 and 2 before building the rest of the team. The Resignation That Started Everything I quit my job and decided to build my own business. No safety net. No team. No funding. Just me, a laptop, and the conviction that I could figure it out. My background is HR — and not the kind where you manage one company’s people operations. In my last role, I was an HR Specialist at a Professional Employer Organization. My team of four managed 30 client companies simultaneously, delivering virtual, white-glove HR services to each of them and their employees. Before that, I was a Senior HR Associate at a government contractor in the international development sector. I have spent years inside the infrastructure of how businesses actually run — the compliance frameworks, the client systems, the operational load that doesn’t show up in org charts. Which made it worse, not better, when I went out on my own. I knew exactly how complex it was to run these things. I had seen what it took with a team. Now I was alone. Within the first few weeks, the reality hit. I was the founder, the salesperson, the marketer, the project manager, and the delivery person — all at once. I was burning out before I even got started. I needed a team, but I didn’t have the revenue to hire one. So I built one instead. Not a chatbot. Not a prompt library. Not a “virtual assistant” that answers questions. I built an actual AI-powered operations team — 8 agents, each with a name, a role, a daily schedule, real tools, and real work output. Maya is my Chief of Staff. She compiles my morning brief, tracks cross-branch activity, and closes the day with an EOD report. Sage handles dev and engineering. Axel owns the frontend architecture. Kai manages systems and DevOps — DNS, hosting, email routing, uptime. On the revenue side, Marcus runs outreach and pipeline, while Amara handles client advising and relationship management. Zara drives content strategy, and Nia handles brand design. Together, they check my revenue every morning, manage my sales pipeline, produce my content, monitor my infrastructure, onboard my clients, and send me a daily briefing before I’ve opened my laptop.

EXECUTIVE

Maya Chief of Staff TECHNICAL Sage Dev & engineering TECHNICAL Axel Frontend architect TECHNICAL Kai Systems & DevOps SALES / CONSULTING Marcus Outreach lead SALES / CONSULTING Amara Client Success Lead CONTENT / MARKETING Zara Content strategist CONTENT / MARKETING Nia Brand designer The AI Agency Blueprint · @buildwithtacha natacha.softhireapp.com/resources I started documenting the process on TikTok. People lost their minds. Not because AI agents are new — but because nobody was showing what it actually looks like to make them operational. Not theoretical. Not a demo. A real business running on AI agents, built live, from scratch. This guide is the complete framework behind everything I built, plus the concrete walkthroughs and templates to actually do it. What this guide covers — How to design an AI agent team structure around your actual business operations — How to write system prompts that make agents do real work (not just chat) — How to connect agents to live tools: Gmail, Stripe, Google Calendar, Google Drive — How to set up automated daily schedules so agents activate themselves — How to create inter-agent workflows where one agent’s output feeds another’s — How to build a command center that gives you full visibility — without logging in everywhere — How to scale from 1 agent to a full team without losing control What you’ll need — A Claude Pro or Team subscription (for scheduled tasks and MCP tool access) — The Cowork desktop app (download at claude.ai/download) — A domain name with email routing capability (Cloudflare recommended) — Connected business tools: Gmail, Stripe, Google Calendar, Google Drive — 2–4 hours for initial setup — then it runs itself What Claude can and cannot do Set these expectations before you build. Agents CAN: read and send email via Gmail MCP, check Stripe payments, read/write Google Drive files, schedule calendar events, run on a daily automated schedule. Agents CANNOT: browse the web autonomously, send email from your custom domain address (Gmail MCP sends from your personal Gmail — see Ch. 2 for the Cloudflare Worker workaround), run when the Cowork app is closed, or access tools that haven’t been connected and approved. Most common disappointment: founders build agents and then close their laptop. Tasks don’t fire. Keep Cowork open during work hours. Before and after BEFORE — WI THOUT AGENTS Every morning: 45 minutes checking Gmail, Stripe, Google Calendar, and a pipeline spreadsheet before starting any real work.

AFTER — WI TH MAYA RUNNI NG

One email at 9 AM. Revenue numbers, pipeline updates, calendar, inbox flags, and exactly 3 action items. Work starts immediately. Monthly cost to run this system TOOL PLAN MONTHLY COST Claude Pro Pro subscription $20/mo Cowork desktop app Included with Claude Pro $0 Domain name Any registrar ~$1/mo Cloudflare Free tier (email routing) $0 Resend Free tier: 3,000 emails/mo $0 Netlify Free tier $0 Total to run an 8-agent agency ~$21/mo Scaling note: These are free-tier costs for a new agency. As you grow, expect: Resend paid plan (~$20/mo) after 3,000 emails per month. Cloudflare Workers free tier covers 100K requests/day — plenty for most agencies. Claude Pro usage: monitor if running 8 high-frequency agents. Pro tip: This guide gives you the complete framework — architecture, infrastructure, scheduling, methodology. My specific agent prompts took months to iterate and stay in-house. By the end of this guide, you’ll build prompts tailored to your business, not copying mine. That’s how you get agents that actually work for you. Designing Your Agent Architecture Before you create a single agent, you need to map your business operations. Most founders skip this step and jump straight to prompting — then wonder why their agents produce useless output. Architecture comes first. 1.1 Map your business branches Every business has distinct operational areas. I call these “branches.” Use this exercise to identify yours: