The Solopreneur’s AI Stack: How I Run 20 AI Agents for $100/Month

Last month, the 20 AI agents I run published 47 articles, optimized SEO on 200+ pages, generated 150 images, produced 3 podcast episodes, and sent 12 email campaigns — the output of a content-and-commerce team. My total software spend for the month was about $100.

Last updated: June 2026

Not $1,000. Not $10,000. About a hundred dollars a month. The point is not the gadgetry — it is what that ~$100/month covers: as of 2026, the delegated labor I used to hire, and the capacity it freed up to run eight sites and an e-commerce business without staffing up. Set against the $8,000-12,000/month a comparable team would cost, the comparison makes itself.

Here is what the swap looks like in plain numbers — the comparison, as of 2026, that drove the whole project:

What I run	Delegated team (before)	Owned AI-agent stack (now)
Monthly cost	$8,000–12,000/month in labor	~$100/month software budget
Who does the work	Virtual assistants and contractors I manage	20 specialized AI agents I own
Coverage	Whatever the headcount could reach	All 8 sites, every day
Oversight	I was the bottleneck for every decision	Confidence-gated: 85%+ auto-executes, the rest queues for my review

I run eight websites, two YouTube channels, and a Japanese tea and kitchenware shop from Hawaii. For a long stretch, the way I kept that running was to hire virtual assistants. This is the behind-the-scenes story of how I replaced most of that delegated labor with an agent stack I own — what it actually cost, what it freed up, and the honest trade-offs.

No hype, no gatekeeping: the real system, the real numbers, and what to weigh first.

What Does ~$100 a Month Actually Replace?

It replaces the delegated labor I used to hire. The agents took over the mechanical work I leaned on virtual assistants for, at a scale that would otherwise mean a team of 8-10 people — the kind of headcount that runs $8,000-12,000/month, which I break down later. The 20-agent stack does it for ~$100/month to run the stack, across every property, every day, without the management overhead.

That is the framing that matters for an operator. The question is not “should I learn AI.” It is “what am I currently paying people and software to do, and how much of it can a system I own do instead?”

The Backstory: From a Delegated Team to an Owned Stack

Three years ago, I did everything by hand.

The visual below shows the complete roster of 20 specialized AI agents across content, SEO, and business operations — the same division of labor you would otherwise hire for.

Every blog post drafted in Google Docs, every image sourced or shot by hand, every product description written one at a time, every meta description typed manually, every internal link placed by scrolling through WordPress and hoping I remembered what I had published.

I was running one website — All Day I Eat Like a Shark, a Japanese cooking and tea blog. One site, and I was already drowning: the content treadmill is real, but consistency at the pace content marketing demands is a full-time job, and I had products to sell, videos to film, and a life to live.

So I did what most operators do: I hired virtual assistants. The results were mixed. Good VAs are expensive; affordable ones need so much oversight you barely save time; training docs go stale the moment you write them. And the core problem remained: I was still the bottleneck for every decision and every “does this match our brand voice” call. The VA model moved the work, but it never removed me from the middle of it.

Then I started building agents — not ChatGPT, which I had used for months to draft, but agents: persistent AI systems with specialized instructions, tool access, and the ability to run multi-step workflows autonomously. The difference between asking ChatGPT to “write a blog post” and deploying an agent that monitors rankings, drafts in your voice, optimizes for SEO, and publishes to WordPress is the distance between a calculator and a spreadsheet.

I built my first agent in early 2025 — a content writer with instructions in a file called CLAUDE.md. Within two weeks I knew this would change what one operator could carry. 18 months later, I run 20 agents across all my websites. They handle 80% of the operational work that used to consume my day, including drafting and optimizing content according to my content strategy for solo operators. Total software budget: ~$100/month to keep running. (It ran about $260/month during the heavy build phase, on the $200/month Claude Max plan; once you are not actively building, the plan steps down and the steady-state cost settles to roughly $100/month.) Here is exactly how it works.

What Is the Three-Layer Stack, and Why Does It Protect Your Margin?

The Three-Layer Stack is what keeps the AI bill low instead of ballooning. It is a routing decision: every task falls into one of three layers, and the layer determines what technology handles it. Get it right and AI is a margin lever; get it wrong and it is another subscription bleeding cash. (Full framework in my AI Content Creation Guide; here is the essential version.)

Layer 1: Zero-Cost Computation

Layer 1 tasks need no AI model — pure computation: scheduled scripts, data aggregation, API calls, monitoring. A central orchestration script, data-pipeline.mjs, coordinates dozens of automated workflows: pulling Google Analytics and Search Console data every Monday, running monthly site health audits, syncing content calendars, monitoring keyword rankings, managing the publishing queue, and backing up everything nightly.

None of this needs intelligence; it needs reliability, and a scheduled Node.js script does it perfectly. Cost: $0 — it runs on hardware I already own.

Layer 2: Near-Zero Bulk LLM

Layer 2 handles tasks that need language understanding but not the best model available. I run Qwen3 32B, an open-source LLM, locally on an NVIDIA RTX 5090. Marginal cost per generation: electricity only, about $15-20/month.

This is the layer that quietly kills the per-seat SaaS bill. Layer 2 handles first drafts across all the sites, FAQ answers for schema (I batch-generated answers for 50 product pages in 20 minutes), product description rewrites — the bulk e-commerce work you would otherwise outsource — plus review sentiment analysis, newsletter summaries, alt text, and meta descriptions.

The same GPU runs ComfyUI with FLUX models for images, Whisper for transcription, and TTS for podcasts.

Cost: ~$18/month in electricity, no API fees. For about 70% of text generation tasks, Qwen3 32B is close enough to a premium model that the difference does not show up in the published output.

Layer 3: Premium Intelligence

Layer 3 is where I spend real money and where it earns its keep — tasks that require the best reasoning: strategic planning, complex analysis, brand-voice matching, quality evaluation, and creative direction.

I use Claude from Anthropic as the Layer 3 brain — specifically Claude Code, the command-line tool that turns Claude into an operating layer for the business. It handles strategic analysis (competitor gaps across 52 domains, quarterly planning), complex content, quality judgment before content ships, agent orchestration, and structural editing.

Cost: ~$72/month — Claude Pro plan (~$25/month, stepped down from the build-phase Max plan) plus ~$47/month in metered API overflow (most inference runs locally on the RTX 5090 at $0 marginal cost, so only the residual volume hits the API). This is the largest line in the software budget.

The key insight for anyone watching their margin: by routing 70% of work to Layer 2 and 20% to Layer 1 — both effectively free — only 10% of the workload hits the expensive model. That 10% is the work that benefits from premium intelligence. The trap is that most operators adopt premium models for everything and overspend on routine work a cheap local model would handle just as well.

Meet the 20 Agents (the Team You Are Not Hiring)

Each agent is a specialized AI worker with its own instructions, tools, and domain expertise — read the roster as a hiring plan you never had to execute. It is the same set of solopreneur tools that runs the operation; for the wider toolkit, see my 17 AI tools.

Content Production:

1. Content Writer — Researches topics and drafts WordPress-ready articles in each site’s templates and brand voice, following recipe, internal-linking, and SEO rules.
2. Content Refresher — Flags stale content via ranking and traffic signals, merges new research into existing articles section-by-section, and republishes with updated dates.
3. Recipe Enhancer — Enriches recipe posts with culinary context, ingredient deep-dives, and cultural background, turning thin pages into comprehensive resources.
4. Repurposer — Turns one blog post into social clips, podcast notes, newsletter excerpts, and YouTube scripts: five formats from one piece. The backbone of my AI content creation workflow, and the content-velocity problem most operators throw a VA at.
5. Newsletter Agent — Writes and schedules email campaigns through Omnisend, pulling from recent content and managing drips.

SEO and Technical:

6. SEO Optimizer — Audits on-page issues, fixes meta descriptions and heading hierarchy, and runs AEO passes for AI Overview and featured-snippet targeting.
7. Schema Injector — Adds FAQ, HowTo, Article, and Recipe structured data. Batch-processed 50 product pages with FAQ schema in one session.
8. Internal Linker — Scans the corpus for contextual linking opportunities and adds 3-5 relevant anchor links per post.
9. Alt Text Agent — Generates descriptive, keyword-aware alt text. It is currently working through a backlog of 7,000+ images across all sites.
10. Site Auditor — Runs monthly health checks for broken links, missing images, thin content, orphan pages, and schema errors.

Research and Analysis:

11. Discovery Agent — Analyzes Google Analytics and Search Console to find keyword opportunities and content gaps for the calendar.
12. Research Agent — Deep dives into topics, competitor strategies, and market trends, producing the briefs that feed the content writer.
13. PAA Pipeline — Harvests “People Also Ask” questions for target keywords, generates answers via Qwen3, and injects them as FAQ sections with schema markup.

Business Operations:

14. Business Advisor — Weekly strategic briefings, revenue analysis, and cross-site performance comparison. My Monday-morning dashboard.
15. Growth Marketer — Analyzes Amazon Ads and Google Ads campaigns, recommends budget shifts, and flags wasted spend.
16. Ads Analyzer — Pulls campaign data, calculates ACoS/ROAS, and generates optimization recommendations.
17. Listing Optimizer — Optimizes Amazon and Walmart listings: titles, bullets, descriptions, backend keywords, and A/B tests. The agent e-commerce sellers value most.
18. Review Bot — Monitors reviews across platforms, performs sentiment analysis, drafts response templates, and alerts on negatives that need attention.
19. Cross-Sell Agent — Analyzes WooCommerce purchase patterns and suggests co-purchase bundles.
20. Cannibalization Resolver — Detects keyword cannibalization, recommends consolidation or differentiation, manages redirects.

Twenty specialized agents, each with clear instructions, boundaries, and quality gates. As the roster diagram above shows, they map onto the content, SEO, research, and operations functions you would otherwise staff — together a content-and-commerce operation that would take a team of 8-10 people to replicate by hand. The Repurposer’s distribution work turns one piece of content into reach across formats — leverage an operator cannot ignore, not vanity output.

How Do Agents Know What to Do? The CLAUDE.md Framework

Every agent is defined by a single file: CLAUDE.md — the operating manual that tells the agent who it is, what it does, how it decides, and when to ask for help. It is the SOP you would write for a human hire, except the agent follows it every time. A well-written CLAUDE.md has six required sections:

1. Mission — what this agent does, in one or two unambiguous sentences.
2. Tools — which APIs, databases, scripts, and services it can access.
3. Decision Framework — how it prioritizes work and handles edge cases.
4. Output Format — exactly what the deliverable looks like (WordPress HTML, JSON, and so on).
5. Error Handling — retry logic, fallback behavior, escalation rules.
6. Quality Gates — the standards output must meet before it ships.

A simplified example — the CLAUDE.md that defines my SEO Optimizer agent:

# SEO Optimizer Agent

## Mission
Apply on-page SEO across all sites: fix technical issues,
optimize meta descriptions, add schema, build internal links,
and run AEO passes for featured-snippet and AI Overview targeting.

## Decision Framework
Check current rank before analysis:
- Rank #1 with snippet -> Skip (already winning)
- Page 1 (#2-10) -> Full SEO + AEO
- Page 2+ -> SEO first, earn rank before heavy AEO

## Quality Gates
- Meta descriptions: 120-155 chars, includes target keyword
- Internal links: 3-5 per post, contextual anchor text
- Confidence threshold: >=0.85 auto-apply, <0.85 flag for review

That confidence threshold is the whole game — it lets agents operate autonomously while keeping a human in the loop for judgment calls. At 85% or more confidence the agent executes automatically; below it, the item is flagged for review.

In practice, about 80% of agent output clears the gate and ships without my involvement. The other 20% lands in my Mattermost review queue, where I approve, reject, or modify it each morning. Without quality gates you get AI slop; with them, a system you can trust — and an answer to the question every operator asks next.

Who Patches It at 2 a.m.? The Maintenance Question

You do, but far less often than you would supervise a team, because the system fails safe rather than fail loud. The confidence-gating that makes the agents useful also bounds the maintenance: anything an agent is unsure about waits in the review queue instead of shipping, so a bad night leaves a few extra items for me in the morning, not bad content live. And because the stack is self-hosted in one place rather than scattered across vendor dashboards, the surface area I babysit is small — the same Layer 1 scripts that run the scheduled work flag when something fails. A deliberate trade: more control and lower cost for owning the uptime, which I would rather do than rent from vendors who each have an outage I cannot fix.

The Hardware Behind the Stack

Running 20 agents requires infrastructure, but not expensive infrastructure. Self-hosted n8n runs on the same server with no SaaS fee; an NVIDIA RTX 5090 GPU server runs Qwen3 32B, ComfyUI + FLUX, and audio locally (a one-time ~$2,000 build, plus ~$15-20/month in electricity); and a Hetzner VPS with a small Proxmox box hosts WordPress for all the sites, dashboards, scraping, and nightly offsite backups (~$28/month combined).

That infrastructure — electricity plus hosting — is already counted inside the ~$100/month software budget above; the only standalone cost is the one-time ~$2,000 GPU. I host it myself rather than on a managed platform because it keeps cost and control on my side. You do not need all of this to start — I built up over 18 months, and your first agent needs just a laptop and a Claude API key.

What Does It Really Cost? The Honest Breakdown

Here is the honest accounting from my own operation, as of 2026 — the numbers I see in my bills. Read the bottom line against the $8,000-12,000/month a comparable team would cost.

Category	Monthly Cost	What It Covers
Claude Pro (stepped down from build-phase Max plan)	~$25	Layer 3 plan subscription, steady-state tier
Metered AI / API overflow	~$47	Residual API volume above the plan; most inference runs locally on the RTX 5090 at $0
Hosting & infrastructure (Hetzner VPS + DB + caching)	~$28	WordPress hosting for all eight sites plus Proxmox monitoring and backups
Local Qwen3 / ComfyUI / n8n	$0	Self-hosted on own hardware; no recurring software fee
Total	~$100/month	Full steady-state operation: eight sites, 20 agents, e-commerce

The steady-state figure is ~$100/month — that covers software, AI, and all hosting and infrastructure together. To replicate the same output with human workers — a part-time content writer, SEO specialist, VA, social media manager, email marketer, designer, and analyst — you would spend $8,000-12,000 per month, conservatively. Set the owned stack next to that team, on the same figures from my operation, and the trade is stark.

Dimension	Owned agent stack	Delegated human team
Monthly operating cost	~$100/month (all-in: software, AI, and infrastructure)	$8,000-12,000
Headcount equivalent	20 agents	8-10 people
My daily operational load	~90 minutes	8-10 hours
Output last month	47 articles, 200+ pages optimized, 150 images, 3 podcasts	Same scope, far higher cost
Who owns it	You (self-hosted)	Vendors and staff

I am not saying AI replaces all human judgment — I still make every strategic decision, film every video, taste every tea, and write the pieces that need my voice (like this one). But for the 80% of operational work that is mechanical, agents do it faster, more consistently, and far cheaper. The real outcome is not replacing yourself, but recovering the margin and the hours a delegated team consumed.

What Does an Operator's Day Look Like Once the Agents Run It?

It looks like directing, not doing. An honest weekday:

Morning (30-45 minutes): I review what the agents did overnight in Mattermost — the content refresher flagged three articles needing updated stats, the SEO optimizer auto-fixed meta descriptions on 12 pages (confidence above 0.85, so no approval needed), the review bot drafted a response to a negative review. I approve two items, modify one, reject one that feels off-brand. Done.

Creative work (2-3 hours): This is my actual job now — filming YouTube videos, writing book chapters, recording podcasts, developing tea blends, photographing products. The work only I can do, instead of writing meta descriptions.

Strategy (1 hour per week): Monday, the business advisor agent generates a briefing — traffic, revenue, ranking changes, pipeline status — and I make a few directional calls. Lean into hojicha this month? Where should the agents focus next?

I am the director, not the worker. Some weeks I spend more on quarterly planning or building a new agent, but the daily operational load has dropped from 8-10 hours to about 90 minutes. The rest goes to the work that builds long-term value.

How Do You Build Your First Agent?

Start with one, not twenty. Pick the single task draining the most of your team's hours — the ROI shows up before you have built anything else.

Step 1: Pick One High-Cost Bottleneck

Choose the task that eats the most hours — or the most outsourced dollars — every week. For most operators it is content drafting (posts, product descriptions, captions), SEO optimization (meta descriptions, schema, internal links), or e-commerce ops (listing updates, review responses, email campaigns). Pick just one.

Step 2: Set Up Claude Code

Claude Code runs Claude as a persistent agent with file access, tool use, and autonomous execution. Install it, get an API key, and run your first session. If you have never used a terminal, it feels intimidating for about 15 minutes — push through; the payoff is large.

Step 3: Write Your CLAUDE.md

Create a CLAUDE.md with the six sections above: Mission, Tools, Decision Framework, Output Format, Error Handling, Quality Gates. A 30-line file that defines the job beats a 300-line one that tries to cover everything.

Step 4: Test With Human Review on Everything

For the first two weeks, set the confidence threshold to 1.0 — nothing auto-publishes. Read every output and refine the CLAUDE.md. This calibration is where the agent goes from generic to genuinely useful, exactly how you would onboard a careful new hire.

Step 5: Increase Autonomy, Then Add Layers

Once you trust the output, lower the threshold so the agent auto-executes where it performs well while you keep review on anything that drifts. When it is reliable, add a local LLM for bulk tasks, schedule it with cron, and build agent number two. The system grows organically — one agent at a time, each reclaiming a few hours or a line item.

The Honest Trade-Offs

I would be lying if I said this was all upside. The real trade-offs to weigh before committing:

Setup time is real. I spent 18 months building this; the first agent took a weekend, the full operation hundreds of hours. An investment that pays compounding dividends — like planting a garden.

Technical skill helps. I am comfortable with the command line and Node.js. If you are not, the curve is steeper — but Claude Code is approachable, and the AI helps you build the system it runs on.

Quality requires vigilance. Agents will confidently produce mediocre work if you let them. The quality gates and confidence thresholds are what separate a useful automation system from a content mill.

Hardware is an upfront cost. The RTX 5090 was $2,000. You can start without it — just a laptop and an API key — but the full three-layer stack needs hardware that pays for itself within two to three months against what you spent on labor and software.

The Bigger Picture

The conversation about AI for business usually splits into two camps: AI will replace everyone, or AI is overhyped. Both are useless to an operator. The practical question is narrower: which line items — labor and software both — can a system you own absorb?

The agents are not replacing me; they are freeing me. The mechanical work consumed the capacity I needed for the work that grows the business. Now it runs in the background while I film a video about brewing hojicha or write a chapter of my next book. That is the promise of an AI stack for operators — not "do less," but spend your time and money on the work only you can do and let a system you own do the rest.

It runs on a ~$100/month budget — 20 agents across eight websites, against the $8,000-12,000/month a comparable team would have cost. If you are an operator drowning in operational work and outsourced overhead, this is the exit — not another VA, not another course, but a system built once and compounding in value. Start with one agent, write your CLAUDE.md, and watch what happens when you stop being the bottleneck.

Want the exact CLAUDE.md configurations I use for all 20 agents? I packaged my complete CLAUDE.md Template Library — all 20 agent configurations, the Three-Layer Stack setup guide, and the confidence-gating framework — into a free download: no opt-in wall, no upsell, just the templates.

---

Frequently Asked Questions

Can an AI agent replace a virtual assistant or a delegated team?

For the mechanical, repeatable work, yes — that is the whole point. The 20 agents took over the researching, drafting, optimizing, formatting, scheduling, monitoring, and reporting I used to hire out, roughly 80% of the daily load. They do not replace strategic judgment, creative work, or relationship-driven tasks; those still need a human. Augment-and-replace the routine work, not eliminate every role.

What is Claude Code, and how is it different from ChatGPT?

Claude Code is a command-line interface that turns Claude (Anthropic's AI model) into a persistent agent with file access, tool integration, and autonomous execution. Unlike ChatGPT's conversational interface, it deploys agents that run multi-step workflows, maintain state, read and write files, call APIs, and operate on schedules without your intervention.

How much does it cost to start building AI agents if I don't have the hardware?

You can start with just a laptop and a Claude API key — $0 upfront, then API usage (typically $10-30/month for a single agent). The GPU and extra servers are optional and come later. Most operators start with one Layer 3 agent on their existing machine and add local hardware only when API costs justify it.

Who maintains the system when something breaks?

I do, but the design keeps that burden small. Confidence-gating means anything an agent is unsure about waits in a review queue instead of shipping, so a failure overnight produces a backlog to clear in the morning, not bad content live. Because the stack is self-hosted in one place, the Layer 1 monitoring scripts catch most issues before they matter.

How long did it take to build this 20-agent system?

18 months total, but not all at once: the first agent took a weekend, and by months 6, 12, and 18 I had 5, 15, and 20. You do not need 18 months to see ROI — most operators report meaningful time savings within 2-4 weeks of their first agent. My timeline reflects continuous optimization, not the minimum viable system.

Pat Tokuyama is the founder of All Day I Eat Like a Shark and Digital Garden Profit. He runs eight websites, two YouTube channels, and a Japanese tea and kitchenware business from Hawaii, powered by 20 AI agents and a stubbornly anti-hustle philosophy. Connect with him on YouTube or at digitalgardenprofit.com.

---