Autonomous AI · Multi-Business Operations
OpenClaw — Autonomous AI Agent Managing Three Businesses
A production-grade autonomous AI agent that runs 24/7 on a VPS, orchestrating operations across three businesses through a single Telegram interface.
3
Businesses
6
Workflows
250+
Posts
24/7
Uptime
20+
Failure patterns
The Problem
I was running three businesses simultaneously — an AI automation agency, a hotel amenities distribution company, and a boutique guesthouse — each with its own marketing channels, ad platforms, and operational workflows. The daily overhead of content creation, social media management, ad monitoring, lead processing, and cross-platform coordination was unsustainable for a solo operator.
I needed something beyond a chatbot. I needed an autonomous agent that could operate across all three businesses with minimal supervision, make intelligent decisions, and know exactly when to act independently vs. when to ask for approval.
The Solution
OpenClaw — a production-grade autonomous AI agent that runs 24/7 on a VPS, orchestrating operations across three businesses through a single Telegram interface.
Architecture
The system is built around three core design documents that define how the agent thinks and operates:
Identity layer — A character document that sets the agent's voice (direct, no-fluff business partner), default operating mode, and recovery behavior (exponential backoff before escalating).
Decision heuristics — An ordered set of principles the agent evaluates against: friction is signal, think before acting, fix it don't report it, self-heal, simplicity first. These create consistent autonomous behavior without hard-coded rules for every scenario.
Operations manual — A 770+ line rulebook covering confidence-based action thresholds, tool selection hierarchies, error handling formats, security protocols, and business-specific alert thresholds.
Confidence Model
What It Manages
1. AutomatAI — AI Automation Agency
Cold email lead generation targeting HVAC and home service SMBs in the US. The agent monitors campaign performance on Instantly.ai, classifies incoming lead replies (interested / not interested / unsubscribe), and drafts follow-up responses for my approval.
2. Dr. Hotellato Zrt — Hotel Amenities Distribution
Exclusive Hungarian distributor of premium hotel amenities. The agent handles content creation (Mon/Wed/Fri) with a multi-step workflow with two human approval gates, cross-posting (daily IG→FB backfill processing 250+ posts), and weekly content calendar (2-week planning cycles).
3. 4CatsShelter — Boutique Guesthouse
A luxury guesthouse in western Hungary. The agent manages Google Ads and Meta Ads monitoring, content creation (Tue/Thu), and performance reporting — with a hard rule to never modify budgets without explicit approval in HUF.
Approval Flow
Key Design Decisions
Human-in-the-loop where it matters.
Content creation workflows have two approval gates — one after concept, one before publishing. Ad budget changes require explicit confirmation with exact amounts. The agent publishes nothing and spends nothing without approval.
Skill-based specialization.
Rather than one generic prompt, the agent loads specialized skills for each task: a copywriter skill with 109-point formatting rules per platform, an image director skill with mandatory product verification protocols, and an integration skill for fetching real product references via browser automation.
Self-healing with institutional memory.
The agent maintains a MISTAKES.md file with 20+ documented failure patterns and their fixes. Before escalating any error, it checks this knowledge base for a pattern match.
Token optimization.
The full workspace (~34KB) fits within a 50K character bootstrap budget. Reference docs live in a subdirectory loaded on-demand rather than every turn, keeping each interaction fast and cost-efficient.
Security Architecture
- • Gateway bound to loopback with token authentication
- • Telegram DM policy on strict allowlist (never allowAll)
- • Credentials isolated in secrets.env, never logged or exposed
- • Token expiry monitoring with 7-day warning alerts
- • Auto-refresh for OAuth tokens; immediate escalation for expired tokens
Results
- ✓ 3 businesses managed through a single Telegram conversation
- ✓ 6 automated workflows running on cron schedules
- ✓ 250+ post backlog being systematically processed with state tracking
- ✓ Zero unauthorized publishes or budget changes
- ✓ 20+ failure patterns documented and auto-resolved before escalation
- ✓ Low marginal cost — optimized token usage across all workflows
What I Learned
Building an autonomous agent that handles real business operations is fundamentally different from building a chatbot. The hard problems aren't in the AI — they're in operational design: knowing when the agent should act vs. ask, building recovery protocols that actually work, and creating institutional memory so the same mistake never happens twice.
The confidence-based action model was the single most impactful design choice. It replaced hundreds of if/else rules with a simple framework the agent applies universally, and it mirrors how a competent employee actually operates.
Tech Stack
| Component | Technology |
|---|---|
| Runtime | Docker Compose on VPS |
| Primary Model | Claude Sonnet 4.6 |
| Fallback Model | Kimi K2.5 (256K context) |
| Image Generation | Gemini 2.5 |
| Memory | Weaviate (vector) + file-based (mistakes, state) |
| Browser | Playwright + headless Chrome |
| APIs | Meta, Google Ads, WordPress, WooCommerce, Instantly, Brave Search |
| Interface | Telegram |
| Scheduling | Cron-based with timezone-aware active hours |
Let's Talk
I take on 3 new clients per month.
The businesses that move first win. Let's find where you're leaving 20+ hours a week on the table.
Book Your Free Audit● 2 spots remain for March — next availability: April