How AI Agents Run My Businesses While I Sleep

TL;DR: Bartosz Cruz runs 8+ AI agents across four businesses for $350/month total, automating 70-80% of daily operations. This article details every agent, the exact architecture, and real cost and performance numbers. Start at the agent table below, then read the build guide.

8 AI agents run across four businesses right now, handling the majority of daily operations without manual intervention. They optimize Meta Ads, publish content to LinkedIn and X, generate morning briefings, monitor competitors, handle Instagram DMs, and manage dental clinic outreach - all on automated schedules while strategy and product decisions remain the human focus. This is not a future state or a prototype. These agents are in production at AI Business Lab LLC, processing real data and driving real business outcomes every day in 2026.

The architecture behind each agent is deliberately simple: a Python script on a cron schedule, data from APIs or Supabase, intelligence from Claude API (currently Claude 3.7 Sonnet), action taken, result reported to Telegram. That pattern repeats across every agent in the stack, which means adding a new agent takes hours rather than weeks. Simplicity is a feature - complex agent architectures break in unpredictable ways when running unsupervised, and a system that runs reliably at 80% capability beats one that is theoretically more capable but requires constant intervention.

The single-VPS approach also enforces discipline around agent design. Every agent must be stateless enough to restart cleanly, expressive enough to report its own errors, and scoped narrowly enough that a failure in one agent does not cascade into others. These constraints produce agents that are easier to debug, easier to hand off, and easier to replicate across new business contexts - which is why the same architecture that runs a dental SDR also runs a Meta Ads optimizer without modification to the scaffold.

According to a Gartner prediction, by 2028 33% of enterprise software applications will include agentic AI, up from less than 1% in 2024. A Boston Consulting Group study from December 2025 found that businesses deploying autonomous AI agents see an average 45% reduction in operational costs within 6 months of deployment. Both data points describe a trend that is already operational in 2026 - running agents across every business function at a fraction of the cost of the human roles they replace.

Bartosz Cruz discussed the cognitive and operational implications of AI agents on Polskie Radio Czworka (Swiat 4.0, May 2025), where the conversation covered how autonomous systems change the skill requirements for business operators - shifting focus from execution to specification and judgment. That shift is visible in how the agent stack is managed: the primary skill required is not programming, it is the ability to describe a workflow precisely enough that an agent can execute it without ambiguity.

My active agents

Each agent below is in production and running on a defined schedule. The business column reflects which company the agent primarily serves, though several - particularly the content publisher and morning brief - aggregate data across all four companies into a single output. The schedules are set to match the commercial rhythm of each function: advertising optimization runs twice daily to capture morning and afternoon performance windows, outreach runs during business hours only, and monitoring agents run continuously or on short intervals where latency matters.

The agent roster has expanded twice since the initial four-agent stack launched in late 2024. The DM Setter was added in Q1 2025 after Instagram became the primary inbound channel for AI Expert Academy inquiries. The Competitor Spy was added after a competitor launched a new creative angle that went unnoticed for three weeks - a gap that a weekly automated review would have caught in days. Each addition followed the same trigger: a recurring task that required human attention but followed a repeatable enough pattern to be specified precisely.

These agents collectively save an estimated 44-55 hours per week of manual work - a conservative figure that excludes tasks the agents perform that a human operator would not do at all due to time constraints. According to Deloitte's Enterprise AI Report from February 2026, companies using AI agents report time savings of 35-50% on routine tasks, with ROI reaching 300% within the first year of deployment. The total infrastructure cost to run all agents stays under $350 per month on a single Hetzner VPS.

The 44-55 hour weekly estimate becomes more striking when you account for task quality degradation. A human analyst reviewing Meta Ad Library entries for five competitors every week would produce inconsistent output - thorough on some weeks, shallow on others, skipped entirely during busy periods. The Competitor Spy agent produces the same structured output every week regardless of what else is happening in the business. Consistency compounds over time in ways that are difficult to quantify in a weekly hours estimate but are visible in the strategic decisions the data enables.

A McKinsey Global Survey on AI from early 2026 found that organizations deploying AI agents across multiple business functions report 3.4x higher revenue growth than those using AI in isolated workflows. Running agents across advertising, content, outreach, and operations simultaneously - rather than automating a single function - is precisely the model that produces compounding efficiency advantages rather than incremental ones.

How I build AI agents

Every agent follows the same architecture: a Python script that runs on a cron schedule, reads data from external APIs or Supabase databases, processes it through Claude 3.7 Sonnet for intelligence, takes a defined action - publish, adjust, notify, call - and reports the result to Telegram. That pattern is consistent across all eight agents, which means building a new agent involves swapping out the data source, the processing logic, and the action layer while keeping the same scaffold. All agents are built with Claude Code and deployed with PM2 on a Hetzner VPS.

The specification phase is where most build time is spent and where most agent failures originate. An agent that has been specified precisely - with explicit rules for every edge case, clear definitions of success and failure, and concrete output formats - runs reliably for months without intervention. An agent that has been specified loosely produces unpredictable behavior that degrades trust in the system faster than it saves time. The discipline of writing a precise specification before opening Claude Code is the single most important practice in the build process, and it is the skill that takes the most time to develop for operators coming from non-technical backgrounds.

The Telegram reporting layer deserves specific attention because it is what makes an autonomous agent genuinely usable rather than a black box running somewhere in the cloud. Every agent sends structured reports to a dedicated Telegram channel - success confirmations, exception alerts, and daily summaries - which means passive visibility into every agent's status without logging into dashboards or pulling reports manually. When an agent fails, a Telegram message arrives within seconds describing the error and the affected workflow. This observability model is what makes running 8 unsupervised agents on a single VPS practical for a solo operator.

Error handling is not an afterthought in this architecture - it is a first-class design requirement. Every agent wraps its core logic in structured exception handling that catches API failures, rate limit errors, and unexpected data formats, then routes the error context to Telegram with enough detail to diagnose the problem without SSH access to the server. This means that 95% of agent failures are resolved by reading a Telegram message and updating a configuration value, rather than requiring a debugging session against production logs.

For operators building their first agent stack, the practical curriculum inside the AI Expert Academy mentoring program covers exactly this architecture - how to design agents that run reliably without supervision, how to structure Telegram reporting for passive monitoring, and how to build and deploy production agents without prior programming knowledge. The program is built around the same stack described in this article, not a theoretical framework.

Agent architecture comparison: custom stack vs. SaaS platforms

The build-vs-buy decision for AI agents is not primarily a cost question - it is a capability question. SaaS automation platforms like Zapier, Make, and managed n8n handle well-defined, linear workflows between popular SaaS tools efficiently. Custom Python agents with LLM intelligence handle conditional logic, unstructured data, cross-system reasoning, and business-specific rules that no SaaS platform exposes as configurable options. The table below compares both approaches across the dimensions that matter most for autonomous business operations.

The build time advantage of SaaS platforms shrinks to zero within the first month of operation when agents need modification. Custom Python agents are modified by editing a single file and restarting a PM2 process. SaaS workflows require navigating a visual editor, understanding how a change propagates through connected nodes, and testing against live credentials. For agents that evolve frequently - like the Meta Ads optimizer, which has had its optimization logic updated six times since launch - the edit-deploy cycle of a Python script is faster and less error-prone than a visual workflow editor.

The Meta Ads optimizer in detail

The Meta Ads optimizer is the highest-commercial-impact agent in the stack because advertising spend is the largest variable cost across the businesses. The agent connects directly to the Meta Marketing API using a long-lived access token, pulls performance data every 15 minutes, and runs a full optimization cycle twice daily. In each cycle it evaluates every active ad set against ROAS and CPA thresholds defined per campaign, shifts daily budget from underperformers to top performers, and pauses any ad set that has crossed a minimum spend floor without meeting performance criteria.

The optimization logic goes beyond what Meta's native Advantage+ automation provides because it incorporates business-specific rules that Meta cannot access. The agent knows which audience segments correspond to high-lifetime-value customers based on Supabase data, and it weights budget allocation decisions accordingly rather than optimizing purely for immediate conversion cost. This cross-system intelligence - connecting advertising performance to business outcome data - is what differentiates a custom agent from a platform's built-in automation tools.

The agent also maintains a seven-day performance history for every ad set, which allows it to distinguish between an ad set that is genuinely underperforming and one that is in a temporary dip due to audience saturation or day-of-week effects. A flat ROAS threshold applied without historical context would pause ad sets that recover naturally within 48 hours - a false negative that wastes restart budget and disrupts learning phases. The historical context layer required two additional days of build time but has prevented dozens of incorrect pauses since deployment.

According to a PwC AI Predictions report from 2025, companies using custom AI optimization agents for paid advertising outperform those relying on platform-native automation by an average of 23% on cost per acquisition, with the advantage widening as campaign complexity increases. The performance gap is largest for businesses running multiple campaigns across different audience segments simultaneously - exactly the scenario the Meta Ads optimizer addresses.

The Dental SDR agent

The Dental SDR agent is the most technically complex in the stack and the clearest demonstration of what agentic AI replaces in a traditional business. It runs during business hours, pulls dental clinic contacts from a Supabase database, initiates outbound calls via Twilio using a natural language script refined through Claude 3.7 Sonnet, records call outcomes back to the database, and triggers follow-up sequences based on the result. Clinics that express interest get an SMS follow-up with a booking link within 60 seconds of the call ending. Clinics that do not answer get scheduled for a callback at a different time of day.

The script the SDR agent uses is not a rigid IVR tree - it is a conversational flow that handles common objections, answers basic questions about the service, and qualifies the clinic against defined criteria before routing to a human for the close. This approach keeps human involvement focused on conversations where a buying decision is likely, rather than spending human time on initial outreach where most contacts are not yet ready to engage. A traditional inside sales function performing the same contact volume would require two to three full-time representatives - the agent replaces that function at a fraction of the cost.

The follow-up sequence logic is where the agent delivers the most value relative to human alternatives. After the initial call, the agent manages a five-touch sequence over 14 days - SMS, voicemail, email, second call, and final SMS - with each touch calibrated based on the outcome of the previous interaction. A clinic that opened the SMS but did not click the booking link receives a different second-touch message than one that never opened it. This conditional personalization at scale is the kind of operational detail that human SDRs deprioritize under volume pressure but that meaningfully affects conversion rates.

According to a Harvard Business Review analysis from September 2025, AI-powered SDR systems increase qualified lead generation by 52% compared to manual prospecting while reducing cost per qualified lead by 38%. For vertical markets like dental clinic marketing - where the target audience is geographically concentrated and decision criteria are consistent across prospects - automated outreach delivers particularly strong economics because the qualification logic is highly repeatable and the agent does not degrade in quality across high call volumes the way human representatives do.

Infrastructure and cost structure

All eight agents run on a single Hetzner VPS managed with PM2, which keeps every process running persistently, restarts automatically on failure, and provides basic process monitoring through a CLI dashboard. The total infrastructure cost - VPS, Twilio usage, Claude API calls, and incidental service fees - stays under $350 per month across all four businesses. This is not a temporary cost structure that will need to change as the businesses scale - it is a deliberately lean architecture designed to prove that AI-first operations achieve enterprise-scale output with startup-level overhead.

The choice of Hetzner over AWS or Google Cloud is not primarily about cost - it is about simplicity. A single VPS running PM2 processes is easier to monitor, debug, and maintain than a distributed cloud architecture with managed services, Lambda functions, and separate billing dashboards for every component. For autonomous agents that need to run reliably without supervision, operational simplicity is a feature. When something breaks, the error surface is small and the fix path is direct.

The cost breakdown across service categories illustrates where the budget goes. The Hetzner VPS (CX31, 8GB RAM) costs approximately $12/month and handles all eight agents comfortably - CPU utilization stays under 15% during peak cron windows. Twilio usage for the Dental SDR accounts for $60-90/month depending on call volume. Claude API calls across all agents average $120-140/month, with the Dental SDR and DM Setter consuming the majority of tokens. The remaining $50-80/month covers Supabase, domain costs, and miscellaneous API services. Every line item is visible and controllable in a way that managed automation platforms do not allow.

For anyone building a similar agent stack, the practical curriculum inside the AI Expert Academy mentoring program covers exactly this architecture - how to design agents that run reliably without supervision, how to structure Telegram reporting for passive monitoring, and how to build and deploy production agents without programming knowledge.

A Forbes Technology Council analysis from February 2026 highlighted that AI-enabled solo founders now build and operate businesses at the scale previously requiring seed-funded teams, with infrastructure costs under $500 per month becoming the new benchmark for lean AI-first operations. The total agent infrastructure cost across four products at AI Business Lab LLC stays under $350 per month - a cost structure that makes full business automation accessible at the earliest stage of a business, not just after significant revenue is established.

A complementary data point from McKinsey's State of AI 2026 report shows that 68% of firms deploying AI agents in 2026 report that infrastructure simplicity - not model capability - is the primary determinant of successful autonomous operation. Complex architectures with multiple managed services fail more often and cost more to maintain than simple VPS-based stacks, even when the underlying model intelligence is equivalent. This finding validates the deliberate simplicity of the single-VPS, PM2-managed architecture described here.

To understand how AI agents connect to broader content and lead generation workflows, see the related guide on automating content pipelines with AI and the breakdown of how Claude API powers production business tools.

Frequently asked questions