[BUG] Excessive token consumption in web search workflows - exponential token usage growth

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?

Bug Description
I've noticed a significant increase in token consumption during relatively simple research tasks.

What previously required a straightforward web search and synthesis is now frequently being expanded into a multi-agent workflow involving multiple research passes, planning stages, validation steps, and intermediate reasoning chains. The result is token usage exploding into the millions for tasks that should be achievable with a fraction of the cost.

For example, a small research request can unexpectedly turn into a 1.5M+ token execution path despite no explicit request for deep research, agent orchestration, or exhaustive analysis.

The depth of reasoning and orchestration should scale proportionally with the complexity of the request.

Environment Info

• Platform: darwin
• Terminal: vscode
• Version: 2.1.160
• Feedback ID: fc88e2b7-162c-48be-be90-92423d93029a

What Should Happen?

Introduce stronger heuristics for determining when multi-agent workflows are actually necessary.
Default to the simplest viable execution strategy.
Require explicit user intent before escalating to deep research or agentic orchestration.
Provide configurable token budgets or reasoning limits.
Allow users to opt out of agentic workflows entirely for routine research tasks.

Error Messages/Logs

Steps to Reproduce

❯ okay now read about how to make a python library of such niche genre appear on top of results on github or google to apply here

⏺ Skill(deep-research)
⎿ Successfully loaded skill
⎿ Allowed by auto mode classifier

⏺ Workflow(dynamic workflow: deep-research)
⎿ Running in background · /workflows to monitor and save

⏺ Research is running in the background across 5 parallel search angles (GitHub topics, PyPI metadata, README SEO, open source discoverability, and
the security/Polars niche). I'll present the findings when it completes — usually a few minutes.

Claude Model

Sonnet (default)

Is this a regression?

No, this never worked

Last Working Version

_No response_

Claude Code Version

2.1.161

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

_No response_