Claude Agent - Sub-Agent Delegation
Core Concept
mcp__plugin_kg_kodegen__claude_agent spawns independent Claude sub-sessions that can execute tasks autonomously. Each agent has its own conversation context, can use tools, and returns a final report. Perfect for parallel research, independent code analysis, or complex multi-step delegations.
Five Actions
SPAWN (Default)
Create a new agent session with initial prompt.
SEND
Send additional prompt to existing agent.
READ
Read current output from agent.
LIST
List all active agent sessions.
KILL
Terminate agent session and cleanup.
Key Parameters
| Parameter | Type | Required | Description |
|---|---|---|---|
action | string | No | SPAWN (default), SEND, READ, LIST, KILL |
agent | number | No | Agent instance (0, 1, 2...), default: 0 |
prompt | string | SPAWN/SEND | Task for the agent to perform |
system_prompt | string | No | Custom system prompt for agent behavior |
await_completion_ms | number | No | Timeout in ms (default: 300000 = 5 min) |
max_turns | number | No | Max conversation turns (default: 10) |
allowed_tools | array | No | Tools agent CAN use (allowlist) |
disallowed_tools | array | No | Tools agent CANNOT use (blocklist) |
cwd | string | No | Working directory for agent |
add_dirs | array | No | Additional context directories |
Usage Examples
Spawn Research Agent
json
{
"action": "SPAWN",
"prompt": "Research all error handling patterns in this codebase. Return a summary of patterns found with file locations.",
"max_turns": 15
}
Parallel Agents for Different Tasks
json
// Agent 0: Research
{
"agent": 0,
"prompt": "Find all API endpoints and document their signatures"
}
// Agent 1: Analysis (concurrent)
{
"agent": 1,
"prompt": "Analyze test coverage and identify untested code paths"
}
Restricted Agent (Read-Only)
json
{
"prompt": "Review this codebase for security vulnerabilities",
"allowed_tools": ["fs_read_file", "fs_search", "fs_list_directory"],
"disallowed_tools": ["terminal", "fs_write_file", "fs_delete_file"]
}
Background Agent with Timeout
json
{
"prompt": "Deep dive into the authentication system architecture",
"await_completion_ms": 60000,
"max_turns": 20
}
Check Agent Progress
json
{"action": "READ", "agent": 0}
List All Agents
json
{"action": "LIST"}
Terminate Agent
json
{"action": "KILL", "agent": 0}
When to Use What
| Scenario | Use Agent? | Why |
|---|---|---|
| Search for keyword in codebase | Yes | Agent explores autonomously |
| Read specific known file | No | Use fs_read_file directly |
| Parallel research tasks | Yes | Spawn multiple agents |
| Write code | No | Do it yourself |
| Complex multi-step analysis | Yes | Agent handles autonomously |
| Simple calculation | No | Overkill |
Best Practices
- •Be specific in prompts - Tell agent exactly what to return
- •Specify output format - Request structured results
- •Use tool restrictions - Limit agent capabilities when appropriate
- •Launch concurrently - Multiple agents in single message for parallelism
- •Trust agent output - Results are generally reliable
Remember
- •Agents are stateless - each invocation is independent
- •Agent results are not visible to user - you must summarize
- •Prompts should be highly detailed - agent works autonomously
- •Launch multiple agents concurrently for parallel work
- •Specify if agent should research only vs write code