Important read on X from - https://x.com/xiangxiang103?s=21 https://x.com/xiangxiang103/status/2041310654158848025?s=46 How to Avoid "Invisible Token Consumption"? Lately, many MAX 20x / 5x users have been reporting: "I'm barely using it, but by Tuesday I've already burned through 50%-77% of my quota""What used to last a week now runs dry in just two days." Anthropic Claude Code team members even posted themselves asking for help debugging, and here's the core takeaway: Most of these "inexplicable depletions" aren't bugs—it's these hidden sources of consumption wreaking havoc. 1. Top Culprit: Subagents (Child Agents) • Every time you spin up a subagent (including Task tools, Plan/Explore agents, custom subtasks), it counts as a brand-new session on its own. • They don't share the main context—each subagent gobbles up tokens independently (easily tens or hundreds of thousands, even millions). • Real-world case: One user racked up 169 subagents in a week, burning through 226 million tokens—averaging a whopping 850,000 tokens per session. • Savings Tips: ◦ Only deploy subagents when you truly need parallel processing or complex breakdowns. ◦ Have subagents use cheaper models (like Haiku) instead of the default Opus. ◦ Once done, promptly /compact or end the session. 2. Second-Biggest Culprit: Automated Background Tasks (Cron / Recurring Scripts) • A "monitoring script" that runs every 5 minutes might seem like just a few thousand tokens per pop, but it explodes over the course of a day. • Deploy watchers, auto-deploy scripts, and custom harnesses (like Hermes, Pi agent, etc.) are the most common traps. • Even if you're not actively doing anything, they're quietly devouring your subscription quota. • Self-Check Method: Run the token analysis script shared by Kieran (GitHub Gist) to instantly see your session counts and subagent consumption breakdowns.