A solo developer in Taiwan built four AI agents that autonomously run his tech agency using Gemini 2.5 Flash's free tier. The system handles content creation, sales lead generation, security scanning, and operations through OpenClaw, an open-source framework running on WSL2 with 25 systemd timers. The agents process about 105 requests daily out of a 1,500-request limit, keeping costs at zero while maintaining 12K+ social media followers.

The token optimization strategy is key to the system's efficiency. Agents read pre-computed intelligence files stored locally (costing zero tokens), then execute single focused prompts with all context injected before acting on responses. Research tasks like RSS feeds, Hacker News API calls, and web scraping use pure HTTP and Jina Reader, avoiding LLM token costs entirely. Creative and analytical work remains the only token-consuming activity.

After an initial $127 billing mistake from using a GCP project instead of AI Studio, the developer refined the system to achieve remarkable results: 27 automated Threads accounts generating 3.3M+ views, 19 intelligence files, and 62 scripts running continuously. The monthly infrastructure cost sits at roughly $5 for Vercel hobby hosting and Firebase. The bilingual site (zh-TW/en) with 21 blog posts demonstrates how AI agents can effectively replace traditional agency functions while maintaining quality through self-review mechanisms and context-aware engagement.