[FEATURE] Enable dynamic sub-agent creation and modification via natural language in dangerously-skip-permissions mode

[leezenn]

Problem

Currently, custom sub-agents must be created and modified through the /agents slash command interface, that makes them registered. This creates friction when working with dynamic agent workflows where agents need to be created, modified, or removed programmatically during active development sessions.

Proposed Solution

Allow natural language creation, modification, and management of sub-agents when dangerously-skip-permissions mode is enabled for project/local scope. This would enable Claude Code to:

• Create new sub-agents through conversational requests
• Modify existing agent capabilities
• Remove obsolete agents
• Automatically register agents without requiring /agents command <---

Benefits

• More intuitive agent management through natural conversation
• Enables dynamic agent workflows for complex development tasks
• Maintains existing security boundaries through project/local scope limitation
• Leverages existing dangerously-skip-permissions pattern for ~~advanced~~ YOLO users

Scope

This feature would only be available when dangerously-skip-permissions is explicitly enabled, ensuring it remains an opt-in capability for users who need dynamic agent management workflows. Or via settings flag, it's up to you.

Notes:

• Make Claude Code to be able to output agent toolbox uppon modification as feedback for the user to be aware of agents landscape and make adjustments on the go.
• At the moment we have to restart claude for the changes to be registered, because otherwise getting: "Error: Agent type 'agent-name' not found." when new agent is manually created or silently won't pick up manually modified instructions. Hence, assuming special registration process.
• Creative ones can only imagine how that would finally unlock meta agents capabilities.