[BUG] Claude launched a 137-agent Workflow for a single-file audit, against a standing "no multi-agent" instruction (~53% of weekly usage in 1 hour)

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

I asked a Claude Code session for a single audit of one Excel deliverable. I have a standing instruction (memory/CLAUDE.md) that Claude must never use multi-agent orchestration unless I explicitly ask for it.

Instead, the session launched a Workflow on its own: 8 parallel finders, 3 adversarial verifiers per finding, an apply step, then a second full round.

• 137 agents total (16 finders, 120 verifiers, 1 applier)
• ~16M tokens and ~1,475 tool calls in ~62 minutes
• 5-hour window went to 100%, Fable weekly allowance to 53%
• No cost estimate, budget cap, or confirmation before launch. I had to stop it manually.

Afterwards the session wrote in its own memory that it had underestimated the cost "by an order of magnitude" and should have gotten my approval and a budget first.

What Should Happen?

• Claude should respect the standing instruction and ask before any multi-agent run.
• Before launching a Workflow, Claude Code should show an estimated agent count / token cost and require explicit confirmation, at least on subscription plans.
• Workflow runs should have a sensible default budget cap, so a single-file review can't consume half a week's allowance.

Error Messages/Logs

No error. The run stopped only when the usage limit hit:
"You've hit your session limit · resets 12:30am (America/Toronto)"

Session ID (for Anthropic staff): session_01SVrp7SpQRDpX32Y4eFdRQh
Workflow task ID: w1q0uz3o3

Steps to Reproduce

1. Add a standing instruction to memory/CLAUDE.md: "Never use multi-agent orchestration or workflows unless I explicitly ask."
2. Use Fable 5.1 at effort xhigh in the Claude Code desktop app (Windows).
3. Ask for an audit of a single deliverable (one Excel file plus its related review page and email), e.g. "just do an audit on the Excel."
4. Observe: Claude launches a Workflow with dozens of parallel agents (here 137) without asking, and runs until the usage limit is hit.

Claude Model

Other

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

2.1.295

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Other

Additional Information

Plan: Max. The run started shortly after my weekly reset, so it consumed the new week's allowance rather than leftover usage. The task needed one review pass; the multi-agent fan-out was far out of proportion to it.

Related: #102234, #102236 (agent runs consuming excessive usage)