Explore agent used unnecessarily for single web page fetch — 84k tokens wasted

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude made incorrect assumptions about my project

What You Asked Claude to Do

When asked to analyze a single public webpage for performance issues, Claude spawned an Explore agent instead of using the WebFetch tool directly. The Explore agent used 84,100 tokens and took 25 minutes to complete a task that WebFetch could have handled in seconds.

What Claude Actually Did

Asked Claude to analyze a webpage for heavy elements causing Elementor publish issues

• Claude launched an Explore agent with a detailed prompt
• The agent made 34 tool calls and consumed 84,100 tokens
• The result was good, but completely disproportionate to the task

Suggested fix:
Add a guard in agent selection logic: if the task involves fetching a single external URL and no local file system access is needed, prefer WebFetch over spawning an Explore agent.

Expected Behavior

Claude should use WebFetch directly for fetching and analyzing a single known URL. The Explore agent is designed for codebase exploration — not web page fetching.

Files Affected

Permission Mode

Accept Edits was OFF (manual approval required)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

1. Open Claude Code in terminal (PowerShell, default permission mode)
2. Ask Claude to analyze a single public webpage for performance issues, e.g.: "Can you read this page and analyze what makes it heavy? https://www.handzhake.com/learn-more/"
3. Observe that Claude spawns an Explore agent instead of using WebFetch directly

Result: Explore agent runs 34 tool calls, consumes 84,100 tokens and takes 25 minutes to complete a task that WebFetch could handle in seconds.

Expected: Claude uses WebFetch directly for a single known public URL with no local file system involvement.

Claude Model

Sonnet

Relevant Conversation

Impact

Medium - Extra work to undo changes

Claude Code Version

v2.1.114

Platform

Anthropic API

Additional Context

_No response_