deep-research workflow spawned 98 agents and consumed 700k+ tokens for a trivial question

Summary

The deep-research skill/workflow was triggered for an extremely simple question: \"How do I set static white RGB on a Razer BlackWidow Chroma V2 at Linux startup without Razer Synapse?\"

The result: 98 agents spawned, 700k+ tokens consumed, before I manually stopped it. The answer was 10 lines of shell commands that Claude already knew without any research.

What happened

1. User asked a simple Linux peripheral configuration question.
2. Claude invoked the deep-research workflow (via the deep-research skill).
3. The workflow fanned out into 98 parallel agents doing web searches, fetching sources, adversarial verification, synthesis — the full pipeline.
4. User noticed the token count and interrupted. Claude cancelled the workflow and answered from existing knowledge in ~10 lines.

Why this is a serious problem

This user is on the Max plan (~€110/month). Token consumption at this scale for a trivial question is not acceptable. The deep-research workflow has no proportionality check — it runs the same heavy pipeline regardless of whether the question is "plan a distributed system from scratch" or "how do I set a keyboard color".

Expected behavior

• The deep-research skill should only be invoked when the question genuinely requires multi-source research that Claude cannot answer from training data.
• There should be a lightweight path: a simple WebSearch call (1-2 searches) for questions that just need a quick fact-check.
• The skill description or guardrails should prevent this overkill for questions with well-known, stable answers.

Impact

• Massive unnecessary token burn on a paid Max plan
• Slow response (workflow runs for minutes)
• No added value — the final answer came entirely from Claude's existing knowledge, not from the 98-agent pipeline

Suggestion

Add a complexity/scope gate before launching deep-research. If the question can be answered from training data with reasonable confidence, answer directly or use a single WebSearch. Reserve the full workflow for genuinely complex, multi-faceted research tasks.

---

Reported on behalf of a frustrated Max plan user.