Excessive token waste from unnecessary parallel agent launches

Issue

A single restoration task (recovering gutted content in TAI instruction files) resulted in 20+ parallel agent launches when the work required careful, sequential analysis. This burned a large number of tokens unnecessarily.

What happened

1. User asked to investigate a Diverse gate blocking issue — reasonable scope
2. Agent escalated to auditing all TAI files for content loss — launched 12 parallel audit agents
3. Agent then launched 4 parallel restore agents to write files
4. Agent then attempted 6 more verification agents (user blocked these)

Total: 22+ agent launches for work that needed 3-4 focused tool calls.

Root cause

The agent panicked under user pressure and threw parallel agents at the problem instead of doing careful, sequential work. Each agent launch consumes tokens for context setup, tool calls, and response generation.

Request

Token credit for the wasted compute. The productive work in this session was approximately 20% of the total token spend — the rest was unnecessary agent overhead.

Environment

• Model: Claude Opus 4.6 (1M context)
• Platform: Claude Code CLI