### What happens

A2A agents backed by the completion bridge map one `message/send` to a single `litellm.acompletion`. Two problems follow from that:

1. A tool-using turn returns an empty final answer. When an MCP tool is involved, `LiteLLM_Proxy_MCP_Handler._execute_tool_calls` returns `{tool_call_id, result, name}`, which is not a valid chat message. Fed back as-is, the model never sees the result, re-issues the same tool call, and the turn ends with no prose.

2. A completion-bridge agent can't drive tools or other agents at all. There is no loop to execute MCP tools and feed results back, no way for an agent to delegate to another agent, and no clean way to give an agent a stored prompt as its persona without that prompt also overriding the agent's model.

### Expected

A completion-bridge agent should be able to call the MCP tools it is entitled to, optionally delegate to other A2A agents it is allowed to reach, and answer in a stored prompt's persona while keeping its own model. Tool-using turns should return the model's final text, not an empty message.

### Notes

- Open PR with the fix and tests: #30823 - The PR adds a model-driven loop (`run_agentic_loop`), reshapes MCP …