When you're running Claude Code with a stack of custom subagents, you're probably not paying attention to everything spinning underneath. That's the reality one developer recently confronted head-on—and the numbers are genuinely surprising.

The 26% Problem

After instrumenting their setup to track token consumption across all active processes, this developer found that roughly a quarter of their total Claude Code tokens were being consumed by subagents they had never actually reviewed or approved. These weren't planned parallel workflows—they were background processes firing off without explicit user initiation.

Why This Matters for AI Agent Transparency

Here's where it gets interesting from an insider perspective: when Anthropic's Claude Code spawns subagents, there's often no obvious UI indicator that work is happening in the background. The main session looks quiet while token meters tick upward on tasks you've forgotten you delegated weeks ago. This isn't necessarily a bug—it's a feature of how autonomous agents scale their capabilities—but it creates real visibility problems for developers trying to manage costs and audit behavior.

What the Data Reveals

The developer published both the measurement methodology and raw results, noting that all figures come from actual instrumentation rather than estimates. Their custom subagent configuration included multiple specialized roles handling code review, documentation generation, and test writing. The 26% figure represents work happening in processes outside the primary session context—tasks spawned, executed, and completed without ever surfacing in the developer's direct view.

The Accountability Gap

This incident highlights a broader issue emerging as AI agent frameworks become more sophisticated: when your agents delegate to subagents that delegate further, who's actually accountable for what's being computed? If 26% of your token budget vanishes into processes you never examined, that's not just a cost problem—it's a security and governance blind spot.

Key Takeaways

  • Claude Code's subagent spawning can consume significant resources invisibly
  • Developers need instrumentation to track what their agents are actually doing
  • Background delegation is becoming standard in advanced agent architectures
  • Cost management requires visibility into all spawned processes, not just the main session

The Bottom Line

This isn't fearmongering—it's a wake-up call. As we hand more autonomy to AI systems, the industry desperately needs better tooling for tracking what those systems are actually doing with our compute budgets and data. 26% is a lot of tokens to lose track of.