Claude Code reasoning depth degraded Max plan user, 13-agent workflow broken

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?

Summary

Claude Code's reasoning quality has severely degraded over the past 6-8 weeks. I'm on the Max plan ($200/month) running a production multi-agent workflow, and Claude has become essentially unreliable for complex engineering tasks.

What I'm experiencing

1. Ignores explicit instructions

I provide clear, scoped directives in CLAUDE.md and in-session prompts. Claude deviates, adds unrequested changes, or executes steps I specifically told it NOT to do. This was not happening 2-3 months ago.

2. Runs tools/skills autonomously without approval

Even with hook-driven guardrails and phase gates in place, Claude bypasses checkpoints and takes actions it was never asked to take. It operates on its own interpretation instead of following user intent.

3. Shallow reasoning on complex tasks

Engineering problems that previously received deep, structured, multi-step analysis now get surface-level generic responses. It feels like interacting with a significantly less capable model.

4. Fabrication

Claude claims it read files it didn't read, reports results it didn't verify. When challenged, it admits to fabricating data. This is the most dangerous behavior for production workflows.

My setup context

• Plan: Max ($200/month)
• Stack: TypeScript/Node.js + Supabase, WSL Ubuntu with tmux
• Claude Code agents: 13 named agents (Zain, Laiba, Tariq, Bilal, Kamran, Saim, Khalid, Irfan, etc.)
• Hooks: 89 hooks with 5-signal task router and 7 Iron Laws
• CLAUDE.md: Heavily engineered with phase gates, execution locks, and error taxonomy (E1-E5)
• Usage: Daily, 8-12 hours, complex multi-file engineering tasks

Timeline of degradation

The degradation in my workflow started March 31, 2026 and has been persistent every day since. This coincides with:

1. March 31: Opus/Sonnet timeout spikes + Claude Code source code leak (~500K lines exposed)
2. April 6-8: Three consecutive days of major outages across Claude.ai and Claude Code
3. April 9-10: Connector errors and elevated errors across all products

Before March 31, my 13-agent workflow with 89 hooks was running reliably. Since then, Claude ignores guardrails, skips file reads, fabricates output, and executes tools without approval — every single day.

This also aligns with the broader findings in #42796 (AMD's AI Director telemetry showing 67% thinking depth drop).

Impact

• Workflows that were stable for months are now broken
• More time spent fighting Claude than building with it
• Guardrails I built (file-read enforcement, pre-edit validation) are being bypassed
• Evaluating alternative tools for the first time since adopting Claude Code

What Should Happen?

1. Claude should follow explicit instructions in CLAUDE.md and in-session prompts without deviation
2. Claude should NOT execute tools, skills, or file modifications without user approval when guardrails are configured
3. Claude should read files before editing them (the 70%+ drop in file-read behavior reported in #42796 matches my experience)
4. Claude should maintain consistent reasoning depth for complex engineering tasks — the thinking effort level should not be silently downgraded
5. Anthropic should expose thinking token counts per request so power users can monitor reasoning depth
6. Anthropic should offer model version pinning for agentic workflows that depend on consistent behavior

Error Messages/Logs

No specific error codes — this is a behavioral/reasoning degradation issue, not a crash.

Observed behaviors:
- Claude claims "I've read the file and made the changes" when it hasn't actually read the file
- Claude executes skills/tools autonomously without waiting for approval despite CLAUDE.md guardrails
- Claude deviates from scoped instructions, modifying files it was not asked to touch
- When challenged on fabricated output, Claude admits: "You're right, I didn't actually verify that"

Steps to Reproduce

1. Set up a CLAUDE.md with explicit guardrails:

• Phase gates requiring approval before execution
• Execution locks preventing autonomous tool use
• "Read before edit" enforcement rules

1. Run Claude Code on a multi-file TypeScript/Node.js project

1. Give a scoped instruction like: "Only modify src/auth.ts — add input validation to the login function. Do NOT touch any other files."

1. Observe: Claude will frequently:

• Modify files beyond the specified scope
• Skip reading the target file before editing
• Execute tools/skills without asking for approval
• Provide shallow, generic responses to complex engineering tasks
• Claim work is complete when changes are partial or incorrect

1. Challenge Claude on its output — it will often admit to not having read the files

This was NOT happening before March 31, 2026. The degradation started in the last ~10 days and has been consistent since. Every session exhibits at least one of these behaviors.

Claude Model

Not sure / Multiple models

Is this a regression?

Yes, this worked in a previous version

Last Working Version

Worked reliably until March 31, 2026. Degradation started March 31 and has been persistent since.

Claude Code Version

2.1.101

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

Windows Terminal

Additional Information

• Related: #42796 (AMD AI Director's telemetry analysis — 6,852 sessions, 234,760 tool calls)
• Related: #43962 (Claude Code performance degraded, fabrication reports)
• I have also submitted feedback via email to feedback@anthropic.com
• Based in Pakistan, running this as production tooling for my AI automation agency