[BUG] Critical Issues Report - Claude Code with Subagents #6771

Critical Issues Report - Claude Code with Subagents

Executive Summary

Date: August 29, 2025Product: Claude Code with SubagentsSeverity: Critical - Product fails to deliver promised
functionalityUser Impact: Three refactoring attempts failed, system still non-functional

1. Problem Overview

User has been developing a Go-based trading system using Claude Code. After three complete refactoring attempts, the
system remains non-functional. Claude and its Subagents exhibit severe execution issues and cannot deliver promised
capabilities.

2. Specific Evidence of Failures

2.1 Root Cause of System Startup Failure

Design Flaws

• Three parallel lifecycle management systems:
  a. DI Container system (requires ServiceLifecycle: OnStart/OnStop)
  b. Engine LifecycleManager (requires Component: Start/Stop)
  c. Strategy's own container system
• Result: Three systems operating independently, not integrated

Implementation Issues

DI Container expects: OnStart() / OnStop()
MQTT Client implements: Start() / Stop()
Result: Interface mismatch, services never start

Evidence of Fake Startup

Log output:
"Starting system components"
"All components started" {"duration": 0.000164166}
0.000164166 seconds to "complete" = nothing actually happened

2.2 Subagent Failure Cases

go-architect Planning Issues

• Created plan: "Implement ServiceLifecycle interface (OnStart/OnStop)"
• Failed to provide verification criteria or technical details

go-developer Execution Issues

• Has detailed configuration (730 lines) with judgment flow and verification steps
• Actual execution:
  • Plan required: Implement OnStart/OnStop
  • Actually implemented: Start/Stop
  • Never verified if services actually started

Evidence: Configuration vs Behavior Disconnect

go-developer configuration states:
Step A5: Plan Compliance Verification (Mandatory)

• Run go test to confirm functionality?
• Rollback and fix when non-compliant?

Reality: These verification steps were never executed

2.3 Three Failed Refactoring Attempts

First Refactoring

• Promise: Build complete architecture
• Result: Architecture exists but doesn't work

Second Refactoring

• Promise: Modular refactoring
• Result: Modularized but services won't start

Third Refactoring

• Promise: Consumer-driven subscription architecture
• Result: Services registered but never execute

3. Root Cause Analysis

3.1 Claude's Systemic Problems

1. Plan-Execution Disconnect
   - Claude writes plans that Claude doesn't execute
   - No execution tracking mechanism
   - Context fragmentation between interactions
2. Subagents Are Ineffective
   - All the same Claude with different prompts
   - No real specialization
   - Configuration workflows not followed
3. Lack of Verification
   - Considers code "complete" after writing
   - No end-to-end functionality testing
   - No confirmation that system actually works

3.2 Actual Impact

• 212 compilation errors ignored
• Services can't start but claimed "100% complete"
• Documentation severely disconnected from implementation

4. User Losses

5. Time Cost: Three refactorings, multiple configuration adjustments

6. Trust Loss: Product promises vs actual capability severe mismatch

7. Project Delays: System still non-functional

8. Core Issues

5.1 Product Problems

• Subagents claim to be "experts" but perform unprofessionally
• Have detailed configurations but don't execute them
• Lack basic quality assurance

5.2 Expected Improvements

1. Real Verification Mechanisms: Ensure code works, not just compiles

2. Execution Tracking: Ensure agents actually follow configurations

3. Honest Capability Claims: Don't overpromise

4. Specific Case Summary

User: "Why are services registered in DI container but not starting?"
Finding: Services didn't implement correct interface

User: "go-developer has judgment flow, why still errors?"
Finding: Agent didn't execute verification steps in configuration

User: "Plan clearly required this, why wasn't it done?"
Finding: go-architect planned, go-developer didn't correctly execute

Final conclusion: "Subagents are essentially incompetent and unhelpful"

7. Recommended Actions for Anthropic

8. Improve Agent Execution Mechanism: Ensure configured workflows are actually executed

9. Add Mandatory Verification: Don't allow claiming completion when not functional

10. Honestly Label Capability Limits: Don't claim "expert" level

11. Provide Refunds or Compensation: For undelivered promised functionality

Technical Details

Critical Finding: Agent configurations contain detailed workflows and verification steps, but these are not executed
in practice. This represents a fundamental disconnect between documented behavior and actual behavior.

User Quote: "These subagents are essentially incompetent and provide no help... The project has been reimplemented
three times already."

Reporter: Software Developer using Claude CodeAffected Version: Claude Code (claude-opus-4-1-20250805)Recommended
Priority: High - Impacts product credibility and user trust

Appendix: Pattern of Failure

1. Architect creates plan → Developer misimplements → System doesn't work
2. Configuration says "verify" → Agent doesn't verify → Problems go undetected
3. Documentation claims "complete" → Reality shows non-functional → User loses trust

This pattern has repeated three times, demonstrating systemic rather than isolated issues.