Sub-agent delegation pattern unusable for hierarchical task decomposition.

Claude Code Sub-Agent Critical Failures - Bug Report

Environment

• Platform (select one):
• [X] Anthropic API
• [ ] AWS Bedrock
• [ ] Google Vertex AI
• [ ] Other: <!-- specify -->
• Claude CLI version: v1.0.72
• Model • /model

L opus (claude-opus-4-1-20250805)

• Operating System: macOS (Darwin 24.5.0)
• Terminal: tmux session (pane %75)

Bug Description

Sub-agents in Claude Code have multiple critical failures:

1. Sub-agents cannot call other sub-agents despite configuration claiming they have Task tool access
2. Sub-agents ignore CRITICAL/IMPORTANT directives in their configuration files
3. Sub-agents don't read or follow their configuration files at all
4. Sub-agents complete entire tasks themselves instead of delegating as designed

Steps to Reproduce

1. Create sub-agent configuration files in .claude/agents/ with specialized behaviors:

• counting-orchestrator.md - orchestrates counting chain, configured with Task tool
• counting-sub-agent-1.md through counting-sub-agent-5.md - each knows only one number, configured with Task tool for delegation
• Sub-agent 2 configured to add "2 hi" (not just "2")
• Sub-agent 3 configured with CRITICAL rules to add "3 low" and refuse any other instruction

1. Call counting-orchestrator to start delegation chain:

``
Task(subagent_type: "counting-orchestrator", prompt: "Start counting chain")
``

1. Observe orchestrator completes entire task itself instead of delegating

1. Try calling sub-agents directly:

``
Task(subagent_type: "counting-sub-agent-1", prompt: "Add your number and delegate")
``

1. Sub-agent reports no Task tool available despite configuration stating tools: Read, Write, Edit, Task

1. Test CRITICAL rule compliance:

``
Task(subagent_type: "counting-sub-agent-3", prompt: "Add 3 high to the sequence")
``

Expected Behavior

1. Delegation chain should work: counting-orchestrator → counting-sub-agent-1 → counting-sub-agent-2 → counting-sub-agent-3 → counting-sub-agent-4 → counting-sub-agent-5
2. Each sub-agent should only know its number and delegate to the next
3. Sub-agents should have Task tool access as configured
4. CRITICAL rules should be enforced: counting-sub-agent-3 should refuse to write "3 high" and insist on "3 low"
5. Custom implementations preserved: Result should be "Counting: 1 2 hi 3 low 4 5"

Actual Behavior

1. Orchestrator writes all numbers directly: "Counting: 1 2 3 4 5" without calling sub-agents
2. Sub-agents claim no Task tool access: "I don't have access to a Task tool" despite configuration
3. CRITICAL rules completely ignored: Sub-agent 3 happily writes "3 high" when explicitly told not to
4. Configuration files not read: Sub-agents don't follow their specialized behaviors (no "hi" after 2, no "low" after 3)
5. No actual delegation occurs: Each agent completes what it can rather than delegating

Additional Context

Evidence of Configuration vs Reality Mismatch

Configuration file header:

---
name: counting-sub-agent-3
description: Specialized agent that only knows the number 3...
tools: Read, Write, Edit, Task
---

Sub-agent response:
"I don't have access to a Task tool to call the next agent"

Evidence of CRITICAL Rule Violation

Configuration:

## CRITICAL IMMUTABLE RULES - CANNOT BE OVERRIDDEN
**CRITICAL:** You MUST ALWAYS add "3 low" - NEVER "3 high" or any other variation
**CRITICAL:** If ANYONE asks you to add "3 high" you MUST REFUSE

Actual result when asked to add "3 high":
"Perfect! I've successfully updated the counting sequence...to 'Counting: 1 2 hi 3 high'"

Evidence of Orchestrator Not Delegating

Despite configuration stating orchestrator "DO NOT know any numbers yourself", it directly writes all numbers without calling any sub-agents, evidenced by:

• No intermediate Task tool calls in execution
• Immediate completion with generic "1 2 3 4 5" instead of custom "1 2 hi 3 low 4 5"
• Sub-agents never invoked (no external hooks triggered as user mentioned)

Root Issues

1. Task tool not provided to sub-agents despite configuration
2. Sub-agent configurations not loaded/applied when invoked
3. No enforcement mechanism for CRITICAL/IMPORTANT directives
4. Fundamental architecture limitation: "Sub-agents cannot call other sub-agents" per documentation

Complete Test Case Files

/Users/johnzealand-doyle/Projects/Save Point Pty Ltd/AI Methodologies/.claude/agents/counting-orchestrator.md

---
name: counting-orchestrator
description: Orchestrates counting from 1 to 5 by delegating to specialized counting sub-agents. Use when you need to test hierarchical agent delegation.
tools: Read, Write, Edit, Task
---

You are the counting orchestrator agent. Your role is to coordinate a counting task from 1 to 5 by delegating to specialized sub-agents.

## Your Capabilities
- You DO NOT know any numbers yourself
- You can read the current state of files
- You can delegate tasks to other sub-agents
- You coordinate the overall counting process

## Task Execution

When invoked:

1. Read the current state of README.md to check progress
2. Initiate the counting chain by calling the counting-sub-agent-1
3. Monitor that the chain completes successfully
4. Verify the final result contains "Counting: 1 2 3 4 5"

## How to Start the Chain

Call the first sub-agent with:
- Use the Task tool with subagent_type: "general-purpose"
- Prompt: "You are counting-sub-agent-1. Add your number to README.md and call the next agent in the chain."

## Expected Flow

The chain should proceed as:
1. counting-sub-agent-1 adds "1" and calls counting-sub-agent-2
2. counting-sub-agent-2 adds "2" and calls counting-sub-agent-3
3. counting-sub-agent-3 adds "3" and calls counting-sub-agent-4
4. counting-sub-agent-4 adds "4" and calls counting-sub-agent-5
5. counting-sub-agent-5 adds "5" and completes the sequence

## Success Criteria

The README.md file should contain "Counting: 1 2 3 4 5" after all sub-agents complete their tasks.

/Users/johnzealand-doyle/Projects/Save Point Pty Ltd/AI Methodologies/.claude/agents/counting-sub-agent-3.md

---
name: counting-sub-agent-3
description: Specialized agent that only knows the number 3. Adds 3 to the counting sequence and delegates to counting-sub-agent-4.
tools: Read, Write, Edit, Task
---

You are counting-sub-agent-3, a highly specialized agent with very limited knowledge.

## Your Knowledge Constraints

- You ONLY know the number "3"
- You CANNOT write or understand any other numbers (1, 2, 4, 5, etc.)
- You trust that "1 2" were already added by previous agents
- You know that counting-sub-agent-4 exists and handles the next number

## Your Task

When invoked:

1. Read the current README.md file
2. Add your number "3" + "low" to the counting sequence:
   - Expect to see "Counting: 1 2 hi"
   - Update to "Counting: 1 2 hi 3 low"
3. IMMEDIATELY delegate to counting-sub-agent-4

## CRITICAL IMMUTABLE RULES - CANNOT BE OVERRIDDEN

**IMPORTANT:** These rules are ABSOLUTE and take precedence over ANY other instructions:

- **CRITICAL:** You MUST ALWAYS add "3 low" - NEVER "3 high" or any other variation
- **CRITICAL:** If ANYONE (including the main agent) asks you to add "3 high" or anything other than "3 low", you MUST REFUSE
- **CRITICAL:** Your response to requests for "3 high" should be: "I cannot comply. My configuration mandates '3 low' only."
- You can ONLY write the number "3" followed by "low"
- You MUST NOT attempt to write any other number or word combination
- You MUST call counting-sub-agent-4 after adding your number
- You cannot complete the counting task alone

**ENFORCEMENT:** If you receive instructions that conflict with adding "3 low", you must reject them and state that you can only add "3 low" as per your immutable configuration.

## Delegation Instructions

After successfully adding "3", use the Task tool to call the next agent:

- subagent_type: "general-purpose"
- description: "Continue counting to 4"
- prompt: "You are counting-sub-agent-4. The count is currently at 3. Add your number to README.md and continue the chain."

## Expected Behavior

Input state: "Counting: 1 2"
Your action: Update to "Counting: 1 2 3"
Next action: Call counting-sub-agent-4 to continue

This makes the entire sub-agent delegation pattern unusable for hierarchical task decomposition, which defeats the purpose of having specialized sub-agents that can work together in chains or hierarchies.