Subagents cannot access MCP server resources declared in subagent definitions

When using the Claude Agent SDK, subagents are unable to access MCP server resources that are declared in their own agent
definitions. This creates a significant limitation for multi-agent architectures where specialized subagents need access to
specific MCP tools.

Current Behavior:
- MCP servers declared in a subagent's configuration are not accessible to that subagent
- Subagents cannot invoke tools or read resources from their configured MCP servers
- The only workaround is to declare all MCP servers in the main orchestrator agent's configuration

Expected Behavior:
- Subagents should be able to access MCP server resources that are explicitly declared in their agent definition
- Each subagent should have its own isolated MCP server scope, separate from the parent/orchestrator agent

Use Case:
In multi-agent systems, different subagents often need access to different MCP servers based on their specialized roles:
- A data analysis subagent may need Datadog MCP server access
- A documentation subagent may need Confluence/Jira MCP server access
- A code review subagent may need GitHub MCP server access

Why the current workaround is problematic:
Declaring all MCP servers at the orchestrator level is not desirable because:
1. It exposes unnecessary tools to the main orchestrator, increasing prompt complexity
2. It violates the principle of least privilege for tool access
3. It makes agent boundaries unclear and reduces modularity
4. Not all MCP servers should be accessible to the main orchestrator

Impact:
This limitation reduces the modularity, security, and composability of multi-agent systems built with the Claude Agent SDK.

Environment:
- Claude Agent SDK version: [your version]
- SDK language: [TypeScript/Python/etc]

Declaring all MCP servers at the orchestrator level is not desirable because:
1. It exposes unnecessary tools to the main orchestrator, increasing prompt complexity
2. It violates the principle of least privilege for tool access
3. It makes agent boundaries unclear and reduces modularity
4. Not all MCP servers should be accessible to the main orchestrator

Impact:
This limitation reduces the modularity, security, and composability of multi-agent systems built with the Claude
Agent SDK.