Everyone asks about the upside of AI agents. Nobody asks what they actually cost. Here's a transparent breakdown of infrastructure, API, and time costs.
I run 50 AI agents across two businesses. People always ask “how much does that cost?” and expect me to say something shocking.
The actual number is less interesting than the breakdown. Because the cost is not where you think it is.
Let me lay this out clearly.
API costs. This is the biggest variable. My agents use Claude’s API as their primary model. Monthly API spend fluctuates between $200 and $400 depending on volume. Heavy months with lots of content generation or research tasks push toward the upper end. Lighter operational months stay lower.
That sounds like a lot until you compare it to the alternative. A single content writer in London costs $3,000 to $5,000 per month. My content pipeline handles work that would require at least one full-time person.
Infrastructure. I run PostgreSQL with pgvector for memory, a MegaRAG instance for document search, and the Ayla OS orchestration layer. All of this runs on two servers. One VPS at around $50/month for the Postiz scheduling and external services. The rest runs on my local machine and a NAS, which I already own.
Total infrastructure cost: roughly $50/month in hosting.
Tooling subscriptions. Postiz for social scheduling (self-hosted, so free beyond the server). PM2 for process management (free). Claude Code for the agent backbone (subscription I would have anyway).
Total monthly cost: $300 to $500. That is the real number.
The exact architecture, memory layers, and delegation patterns I use to run 50 agents across two businesses.
Get the AI Agent Blueprint →Here is where the transparency gets uncomfortable.
Development time. Building Ayla OS was not a weekend project. I have invested hundreds of hours over several months designing the architecture, building the memory system, writing the cron infrastructure, debugging agent failures, and iterating on prompts.
If I billed my own time at my consulting rate, the development cost would be significant. This is the honest truth about custom AI systems. The software is cheap. The expertise to build it is not.
Maintenance. Things break. An API changes. A prompt that worked fine for weeks suddenly produces garbage because the model got updated. A cron job silently fails and nobody notices until the output stops appearing.
I spend roughly 3 to 5 hours per week on maintenance across all 50 agents. Some weeks it is zero. Some weeks a cascade failure eats an entire afternoon.
Prompt iteration. This one is sneaky. Getting an agent to produce consistently good output requires testing, tweaking, and retesting. The first version of any agent is mediocre. Version five is usually where it starts delivering real value.
Each iteration takes time. Multiply that across 50 agents and you have a serious time investment.
Here is how I think about whether this is worth it.
My 50 agents collectively save approximately 120 to 150 hours of human work per month. That includes content creation, research, operations, reporting, and monitoring tasks that would otherwise require people.
At even a modest $30/hour valuation, that is $3,600 to $4,500 in recovered time per month. Against a $300 to $500 direct cost, the return is 7x to 15x.
But that is only half the equation. The other half is what I do with the recovered time. If I use those hours to take on one additional consulting client per month, the revenue impact dwarfs the cost comparison entirely.
The agents do not just save money. They create capacity.
If you are thinking about building your own agent system, here is where to allocate your budget.
1. API access (60% of direct costs). This is unavoidable. Good models cost money. Cheap models produce cheap output. I have tried to cut corners here and it always backfires.
2. Infrastructure (20%). Keep this minimal. You do not need expensive cloud deployments early on. A VPS and a local machine handle 50 agents fine.
3. Monitoring and alerting (10%). Invest early in knowing when things break. A silent failure is worse than a loud one. I run health checks every hour and stall detectors every two hours.
4. Documentation and SOPs (10%). Write down how everything works. Future you will thank present you. When an agent breaks at midnight and you cannot remember how the memory routing works, documentation saves hours.
Do not compare the cost of 50 agents to zero agents. Compare it to the alternative way of getting the same work done.
Option A: Hire three people to handle content, operations, and research. Cost in London: $12,000 to $15,000/month minimum. Plus management time, office space, benefits.
Option B: Run 50 agents. Cost: $500/month plus your own time for maintenance and development.
Option B does not replace the humans entirely. I still have a team of six. But those six people do higher-value work because the agents handle the repetitive volume.
If I were starting over with what I know now, I would change three things.
Start with fewer agents. I built some agents that do not justify their maintenance cost. Three agents that I barely use still need occasional attention. Every agent is a commitment.
Invest in monitoring earlier. I built health checks and stall detectors months after they should have existed. Several failures went unnoticed for days.
Budget for prompt iteration. I underestimated how long it takes to get an agent from “it works” to “it works well.” Plan for three to five iterations per agent. Budget the time accordingly.
You do not need 50 agents and you do not need $500/month.
One agent on Claude’s API with a well-written prompt costs pennies per day. Start there. Prove the value. Then scale.
The newsletter covers the actual cost breakdowns as I add new agents, including what works and what I am cutting.
Transparency about costs is rare in the AI space. Most people want to sell you the dream without showing the receipt.
Subscribe to The Signal
One issue per week. The biggest AI build of the week, the top five videos worth your time, and an operator's read on what actually matters. For builders who ship, not dreamers who scroll.
No spam. Unsubscribe anytime. Just The Signal.