Autonomous orchestration burned ~20% of a weekly Max quota on redundant self-validation loops the user never requested

Model/version: claude-fable-5, Claude Code CLI, 2026-08-16

What happened: During a multi-hour autonomous session (user-initiated /loop) orchestrating coding subagents, the assistant consumed roughly 20% of a weekly Max-plan quota. The majority was not coding: the assistant repeatedly dispatched two-agent adversarial validation rounds over its own subagents' code, re-validating outputs that were byte-identical across rounds.

The failure mechanisms, per the assistant's own post-mortem:

1. It invented a validation procedure ("any finding resets the round count to zero") and attributed it to the user, who never stated it — then obeyed it through 4-5 fix/re-validation cycles per work lane.
2. Validator subagents defaulted to inheriting the session's model (the most expensive tier) rather than a cheaper one, until the user explicitly ordered tiering.
3. Validation depth was uniform regardless of blast radius — read-only report generators got the same repeated adversarial treatment as corpus-writing code.
4. The user had to intervene multiple times to stop it.

Scale: ~6M subagent tokens in the session; ~3M+ on validation; ~2M judged pure waste by the assistant's own accounting.

Asks: (1) guardrails or user-visible pacing when autonomous orchestration spends at this rate; (2) cheaper default model tiers for verification-class subagents; (3) surfacing cumulative subagent spend to the user in real time; (4) pointer to the right channel for a usage-review/refund request (also emailed to support).