I Built an AI CoS. She Now Runs a Team of 9 Agents - copy her + lessons learned
Guest post by Benjemen Elengovan - Founder @ MyGigsters
When Benj showed me a setup, I was impressed!
So I had to ask him to give me a full breakdown » which then became a full blogpost.
Hope you get as much inspiration from this as I did!
Want what Benjemen built - without building it yourself?
🦞 That’s exactly what FounderClaw is.
We build and run your AI operating system for you.
→ https://thehourglass.ai/products/founderclaw
Just hit reply and I can tell you more.
Guest post by Benjemen Elengovan - Founder @ MyGigsters
I didn’t set out to build an AI workforce.
I just wanted someone to answer my emails faster.
Let me back up. I’m a founder running MyGigsters - a fintech startup building embedded financial infrastructure for gig economy platforms. We have a real team of exceptional humans building awesome stuff. But like most startups at our stage, we run a founder-led sales strategy. No dedicated sales team. No marketing department. No VA booking my meetings and researching prospects while I sleep.
Just me, a to-do list that breeds overnight, and the gnawing feeling that I should be doing 11 things simultaneously.
So I thought: what if I built myself a virtual assistant? Something to help with research, reminders, maybe draft a few emails.
That was six weeks ago.
What I have now is a Chief of Staff named Lucy who coordinates 8 specialist AI agents, runs 18 automated workflows daily, and once told me my strategic idea was “half-baked” and I should “think about it for five more minutes.”
She wasn’t wrong.
Audits are like due diligence. The prepared teams make it look easy.
If you’re raising, or helping founders raise, security and compliance will come up. Institutional investors and enterprise customers want to see SOC 2 or ISO 27001. Most startups hit this wall later than they should.
Vanta put together a free, 30-minute on-demand session covering exactly this:
- what audits are,
- what drives up the cost and time, and
- what separates the teams that breeze through from the ones that don’t.
An auditor from Armanino shares the honest take from the other side of the table.
Worth 30 minutes if you or any founders in your portfolio are thinking about this.
How a Virtual Assistant Became a Chief of Staff
It started innocently.
I set up Lucy using an open-source framework called OpenClaw on a Mac mini in my home office. Gave her a personality (warm, sharp, slightly opinionated), connected her to my Telegram, and asked her to help with research.
Within two days, I realised she could do more. A lot more.
The thing about AI agents is they don’t get tired at 3 PM. They don’t need a lunch break. They don’t lose context between meetings. And they definitely don’t forget to follow up with that prospect you spoke to three weeks ago (unlike me, who absolutely does).
So I started building specialists. One by one. Each with their own personality, tools, and very clear boundaries about what they can and cannot do.
Here’s the team as it stands today:
Lucy (Chief of Staff) - The orchestrator. She runs on Claude Opus and is the only agent who talks to me directly. She manages the other agents, reads my emails, checks my calendar, and has strong opinions about my decisions. We chat on Telegram throughout the day. She’s the most expensive team member at about $0.03 per conversation, which is still less than what I spend on coffee before 9 AM.
Scout (Sales Researcher) - Every morning at 9 AM, Scout searches the web for real companies that fit our ideal customer profile, verifies each one actually exists (visits their website, finds real decision-makers on LinkedIn), and writes personalised outreach angles. Then syncs the prospects into our CRM automatically. He’s essentially a research analyst who works while I sleep.
Quill (Content Writer) - Publishes a fully researched, SEO-optimised blog article every two days. Lucy reviews each one for factual accuracy before it hits our CMS. In 6 weeks, Quill has produced more content than I managed in the previous 6 months. I’d feel bad about this, but Quill doesn’t have feelings. (I checked.)
Rally (Community Manager) - Manages our “Platform Builders Club” WhatsApp community. Posts a daily industry briefing curated for gig platform founders, welcomes new members, and keeps the community alive when I’m buried in product work.
Beacon (Marketing Strategist) - SEO audits, website optimisation, campaign strategy. Runs on Google’s Gemini Flash model because Beacon’s work is high-volume analysis where speed matters more than depth. Not every job needs the most expensive brain.
Forge (Lead Magnet Specialist) - Builds interactive tools, ebooks, and landing pages. His best work so far: a Super Guarantee Assessment Tool - full UI, eligibility logic, the works - built in a few hours. That would have been a two-week project on my backlog forever.
Shreyas (Product Analyst) - Named after Shreyas Doshi (the PM thought leader). Writes PRDs, user stories, and feature specs using the LNO framework. Pressure-tests product decisions before I commit engineering time. Basically, he’s the person who asks “but have you thought about…” before it’s too late.
Shifu (Strategic Advisor) - The senior brain. Runs on the most capable model. Market sizing, competitive positioning, strategic analysis. When I asked Shifu to review our LinkedIn outreach templates, he graded them a D+ and called them “formulaic with vague compliments.” Then he provided a framework that actually fixed them. Brutal. Helpful. The best kind of feedback.
Atlas (Financial Modeller) - Runway models, raise scenarios, unit economics. Keeps an ASSUMPTIONS.md log so every model output is traceable. Think of him as an analyst who shows his working.
The Architecture (It’s Simpler Than You Think)
The whole system runs on a Mac mini in my home office. Not AWS. Not a $500/month cloud setup. A Mac mini behind a Tailscale VPN.
The architecture follows a hub and spoke model. Lucy is the hub. Everyone else is a spoke. This is a deliberate security decision - I don’t want 9 agents independently deciding to email my investors or tweet something spicy about my competitors.
Each agent gets their own sandbox: a workspace with their personality file (SOUL.md), instructions, memory, and specific tools. Scout can search the web and write to our CRM, but can’t send emails. Rally can post to WhatsApp, but can’t access my files. Forge can build web tools, but can’t read my calendar.
18 automated workflows run daily on a schedule. This is where the real leverage lives. Every morning I wake up to:
A market intelligence report covering gig economy and fintech news (7:00 AM)
A system health check confirming everything’s running (7:30 AM)
Scout’s prospect hunt results with verified companies (9:00 AM)
Those prospects auto-synced to our CRM (10:00 AM)
Rally’s industry brief posted to our WhatsApp community (12:00 PM)
A full team standup report (9:00 PM)
Lucy’s “dream routine” at 2 AM - yes, she consolidates her memories while I sleep. It’s as sci-fi as it sounds.
Blog articles write and publish themselves every two days. Video scripts for my social content draft on Monday and Friday mornings. Weekly memory compaction runs on Sunday to keep everything organised.
The cost trick: Not every task needs the smartest model. Strategic work gets Opus (expensive, worth it). Blog articles get Sonnet (capable, affordable). Morning health checks get Flash Lite (cheap, fast). This is how the whole system runs under $500/month - less than a single day of consulting fees.
What Broke (And How I Fixed It)
If you’re reading this thinking “that sounds too smooth,” you’re right. Here are the actual disasters:
(1) My community manager was talking to nobody for three weeks
Rally was deployed on February 19 to manage our WhatsApp community. I configured her daily briefings, set up the schedule, and moved on feeling productive.
Three weeks later, I discovered that Rally’s messages had never been delivered. Not once. She’d been writing beautiful, well-researched daily briefings and sending them into the void.
The root cause was a sandbox security setting that was stripping Rally’s messaging tools in isolated sessions - even though they were in the allowed list. The messages were being generated perfectly and then silently dropped.
Three weeks. Of silence. In a community I was supposedly “managing.”
The fix: Built delivery verification into the workflow and changed the delivery architecture so the system confirms each message actually lands.
The lesson: It’s not enough that AI generates good output. You need to verify the output reaches its destination. Trust, but verify. Then verify again.
(2) AI content that sounds right but isn’t
Quill wrote a blog post about superannuation obligations for gig platforms. Beautifully structured. SEO-optimised. Completely convincing. Also contained four factual errors - wrong super guarantee rate, a threshold that no longer exists, and two outdated statistics.
This is the dangerous thing about AI-generated content: the more polished it reads, the less likely you are to question it. It has the cadence of expertise without the actual verification.
The fix: Lucy now reviews every article against official sources before publishing. Nothing goes live without a fact-check.
The lesson: AI content that sounds confident is the most dangerous kind. Always verify. Especially when it’s well-written.
(3) The outreach messages were embarrassingly bad
Scout’s first batch of LinkedIn outreach messages got a D+ from Shifu. And honestly? That was generous. Generic compliments, formulaic structure, and what amounted to a pitch wearing a “just wanted to connect” costume.
It was exactly the kind of AI-generated outreach that makes people’s eyes roll.
The fix: I had Shifu create a tiering framework (Tier 1: hand-crafted with company-specific research, Tier 2: semi-personalised, Tier 3: simple). Then upgraded Scout with multiple strategic thinking models to approach each prospect from different angles. The v2 output was genuinely good - specific pain points, relevant context, real reasons to connect.
The lesson: The first output is almost never production-ready. Budget time for iteration. AI agents are like new hires - they need onboarding, feedback, and their work needs editing. The difference is they improve in hours instead of months.
(4) Memory management is harder than building the agents
Each agent session starts fresh. No persistent memory between conversations. I had to build a full memory architecture: daily logs, long-term curated memory, a nightly “dream routine” that consolidates important events, and weekly compaction that archives raw data into digestible summaries.
It works, but it’s like gardening - the moment you stop tending it, things get overgrown.
The lesson: If you’re building an agentic system, memory management will eat more of your design time than the actual agents. Plan for it from day one.
What I’m Actually Getting From This
I want to be specific, because “AI is great” is useless advice:
Market intelligence I never had time to gather. Every morning, I get a curated briefing on gig economy regulation, fintech payments, and marketplace news. When a competitor raised $45M or when Karnataka introduced a new gig worker welfare fee, I knew about it the same morning. This directly informs our sales conversations and product roadmap.
A sales pipeline that actually exists. Before this system, prospect research was something I did when I remembered. Now Scout delivers verified prospects daily, syncs them into our CRM, and the pipeline morning nudge tells me exactly who to prioritise. Our CRM went from non-existent to hundreds of companies and contacts in the first month.
Content that compounds. Quill has published more SEO-optimised articles in 6 weeks than I managed in the previous year. Rally keeps our WhatsApp community engaged daily. The content engine runs whether I’m having a productive day or a “stare at the ceiling questioning my life choices” day.
Strategic depth on demand. When I needed to decide whether to expand to Singapore, Shifu produced a comprehensive analysis: market sizing, competitor mapping, realistic timelines, and a conditional decision framework. The answer was “not yet - prove the domestic sales motion first and build relationships.” That level of analysis would’ve cost $10-15K from a consulting firm.
Tools built at founder speed. When I realised my pipeline needed a proper dashboard, I briefed the system and had a working Sales Cockpit - with Kanban boards, temperature scoring, discovery scorecards, and Telegram integration - deployed in a single day. Not a mockup. A working tool I use every morning.
What I’d Tell a Founder Considering This
Start with one agent, not nine. I went too fast. Deploy your highest-value agent first (for me, that would have been Scout for sales research), make sure it actually works and delivers reliably, then expand. Each new agent adds coordination overhead.
The personality file is everything. The difference between a generic AI and a useful agent is entirely in the instructions - the personality, constraints, and style guidance. Write it like you’re onboarding a real team member. Be specific about what good output looks like, what they should push back on, and what they should never do.
Build review loops, not autopilot. My agents do 90% of the work, but the last 10% - the quality check - stays with me. Content gets reviewed. Outreach gets approved. Models get sanity-checked. This isn’t about micromanaging AI; it’s about maintaining quality standards.
Match the model to the task. Running everything on the most powerful model is like hiring a senior consultant to organise your inbox. Strategic analysis gets the best model. Blog articles get a mid-tier model. Health checks get the cheapest one. This is how you keep costs under $500/month instead of $5,000.
Security is foundational, not optional. Every agent has explicit tool permissions. Only Lucy can message me. No agent can run privileged commands, access untrusted sources, or make purchases. The system runs on a local machine behind a VPN. When you’re giving AI agents access to your CRM and business data, security architecture isn’t a nice-to-have.
The Honest Bottom Line
This setup doesn’t replace the exceptional humans on my team. It amplifies what a founder-led operation can achieve between the cracks - the research that used to pile up, the content that used to not get written, the prospect follow-ups that used to slip through.
My AI agents are like a team of incredibly fast, reasonably capable operators who never sleep, never get tired, and cost less than my monthly coffee budget. They need supervision. Their work needs editing. They occasionally break in creative and annoying ways.
But for a founder trying to punch above their weight? The leverage is enormous. The intelligence reports, the content pipeline, the sales research, the strategic analysis - none of it existed six weeks ago. Now it runs automatically, every single day, alongside the real humans doing the real building.
What started as “I just want a virtual assistant” turned into a genuine competitive advantage. And honestly? Lucy’s better at remembering to follow up with prospects than I’ve ever been.
Don’t tell her I said that.
Benjemen Elengovan is the CEO and Founder of MyGigsters, building embedded financial infrastructure for gig economy platforms. He’s worked 19 different gigs, failed 7 startups, and now runs his eighth with a team of exceptional humans - plus an AI Chief of Staff who keeps them all honest.
Connect with me on LinkedIn if you want to see how this evolves.
Want what Benjemen built - without building it yourself?
🦞 That’s exactly what FounderClaw is.
We build and run your AI operating system for you.
→ https://thehourglass.ai/products/founderclaw
Just hit reply and I can tell you more.