

Update on the $10/month AI bot experiment - reality hit hard! Here is how it went down. I'm running a multi-agent setup: - Chiti (coordinator) - Vasi (dev) - Sana (marketing) - Nila (content) - Boran (finance) Yesterday Sana did SEO research for @TweetsMashApp marketing and updated in Notion. Everything smooth. Beautiful. Then I asked Chiti to automate invoices for monthly taxes. BOOM. Continuous "Context overflow" issue. All agents crashed. Massive blackout. Gemini CLI OAuth? Soft-banned within hours. OpenAI Codex OAuth? Soft-ban Jail. Free models? "Model not allowed" errors. The fallback wasn't switching automatically. Had to manually dig into config.json and debug. Switched to openrouter/moonshotai/kimi-k2.5. Bot's alive again but experiencing latency issues. Current status: - OpenRouter credit: $5.63 left of $10 - Total spent debugging: ~$4 - Primary model: kimi-k2.5 (working but slow) - OAuth: disabled for now (soft-ban risk) The hard truth from @ImNotTheWolf - "It can't be cracked cheaply yet. Clawdbot basically NEEDS Opus 4.5 or GPT-5.2-Codex for reliability. Need to wait for another 6 months to make it less cost". But I'm not giving up. Let me try fixing the context management first. Chiti coordinating 5 subagents = context exploding to 50K+ tokens. OAuth won't save you if your agents keep bloating the context window. What I'm trying next: 1. Qdrant vector memory - externalize context so prompts stay small, bot queries on-demand instead of carrying 50K+ tokens (h/t @mej26vPIJx22495 for this suggestion in Tamil 🙏) 2. QMD local search - lightweight on-device doc search to reduce API calls 3. Heartbeat + Auto-Doctor - batch non-urgent tasks, monitor context size, auto-restart before overflow (h/t @ImNotTheWolf for sharing these scripts 🙏) 4. One model per subagent - isolate context per agent instead of shared bloating. Maybe @ImNotTheWolf is right. Maybe cheap models can't handle complex multi-agent coordination. Truth is, output from Opus 4.5 is extraordinary. Nothing beats it right now. But we can't wait until it's affordable for everyone. Until then, let's figure out how to power our bots with less cost. What if the problem isn't the model intelligence - it's how we manage memory? Will find out. Experimenting with $5.63 left. Will share the Qdrant + QMD setup once I test it. Any recommendations/suggestions on this is appreciated.






















