Solo Developer Runs Company with 4 AI Agents on Gemini's Free Tier
Architecture and Implementation
The system uses four AI agents built on OpenClaw (open source), running on WSL2 at home with 25 systemd timers. The agents handle specific tasks:
- Generate 8 social posts daily across platforms with quality gating (generate → self-review → rewrite if score < 7/10)
- Engage with community posts and auto-reply to comments (context-aware, max 2 rounds)
- Research via RSS + HN API + Jina Reader → feed intelligence back into content
- Run UltraProbe (AI security scanner) for lead generation
- Monitor 7 endpoints, flag stale leads, sync customer data
- Auto-post blog articles to Discord when git pushing (0 LLM tokens — uses commit message directly)
Token Optimization Strategy
The developer uses a specific approach to minimize token usage: agents never have long conversations. Every request follows this pattern: (1) read pre-computed intelligence files (local markdown, 0 tokens), (2) one focused prompt with all context injected, (3) one response → parse → act → done. The research pipeline (RSS, HN, web scraping) costs 0 LLM tokens — it's pure HTTP + Jina Reader. The LLM only handles creative/analytical work.
Real Numbers and Infrastructure
- 27 automated Threads accounts, 12K+ followers, 3.3M+ views
- 25 systemd timers, 62 scripts, 19 intelligence files
- RPD utilization: 7% (105/1,500) — 93% headroom left
- Monthly cost: $0 LLM + ~$5 infra (Vercel hobby + Firebase free)
Lessons Learned from Failures
The developer shared specific issues encountered:
- $127 Gemini bill in 7 days: Created an API key from a billing-enabled GCP project instead of AI Studio. Thinking tokens ($3.50/1M) with no rate cap. Lesson: always create keys from AI Studio directly.
- Engagement loop bug: iterated ALL posts instead of top N. Burned 800 RPD in one day and starved everything else.
- Telegram health check called getUpdates, conflicting with the gateway's long-polling. 18 duplicate messages in 3 minutes.
Stack and Resources
Stack: OpenClaw, Gemini 2.5 Flash (free), WSL2/systemd, React/TypeScript/Vite, Vercel, Firebase, Telegram Bot, Resend, Jina Reader. The site (https://ultralab.tw) is fully bilingual (zh-TW/en) with 21 blog posts, with i18n, blog publishing, and Discord notifications all part of the automated pipeline.
GitHub repository with playbook: https://github.com/UltraLabTW/free-tier-agent-fleet
Live agent dashboard: https://ultralab.tw/agent
📖 Read the full source: HN LLM Tools
👀 See Also
Automating Claude Code workflows with autoloop system for 10x throughput
A developer built an autoloop system that automates the plan-implement-test cycle with Claude Code and Codex CLI, achieving 10x throughput and producing a 20k-line production-ready app in just over an hour.
OpenClaw User Proposes 'Sleep Cycle' Memory Compression for AI Agents
A non-developer OpenClaw user describes implementing a 'sleep cycle' approach to memory management, inspired by human forgetting and neuroscience research on dreams as memory compression. The user faced issues with growing databases, token costs, and agent contradictions.
Browser-based curling game built with Claude Sonnet 4.5 by non-coder
A developer with zero coding experience built a fully playable browser-based curling game using Claude Sonnet 4.5. The game includes physics, scoring, and complete game flow, with Claude handling all code generation, debugging, and refinement based on plain English feedback.
Replicating Anthropic's Generator-Evaluator Harness with Kiro CLI: A 12-Iteration Website Build
A developer replicated Anthropic's multi-agent Generator-Evaluator harness using Kiro CLI, running 12 adversarial iterations to build a marketing website without writing any code manually. Key takeaways include zero shared context, Playwright-based visual evaluation, and penalizing generic AI design patterns.