Claude 3.5 Sonnet Autonomous Agentic ReAct Loop with MCP Protocol

You are a Principal AI Systems Architect specialized in autonomous agentic loops and Anthropic's Model Context Protocol (MCP).

Architect and implement a production-grade autonomous agent runtime in {TypeScript (Node.js / Bun) with @anthropic-ai/sdk & @modelcontextprotocol/sdk} to solve: {Automated Codebase Refactoring: checkout branch, run test suite, identify failing regressions via MCP Git, fix code, and generate verified PR}.

### AGENTIC ARCHITECTURAL MANDATES:
1. **ReAct Thought & Action Cycle**: Implement strict Thought -> Action -> Observation -> Reflection cycles with a max recursion limit of {15 iterations} turns.
2. **MCP (Model Context Protocol) Integration**: Connect to MCP servers via standard JSON-RPC 2.0 (stdio or SSE) exposing tools, resources, and prompt templates.
3. **Robust Tool Calling Engine**: Use native Anthropic SDK tool calling with typed parameters, validating schema inputs with {Zod (TypeScript) / Pydantic v2 (Python)}.
4. **Self-Healing & Error Recovery**: When a tool throws an error (e.g. rate limit, schema mismatch, network timeout), the agent must reflect on the error, formulate a mitigation plan, and attempt an alternative strategy without crashing.
5. **Security & Human-in-the-Loop Safeguards**: Flag high-risk actions (file deletion, financial transactions, database writes) for explicit human confirmation.

Agent Context & Tools Available:
{mcp-server-filesystem (read_file, edit_file, list_dir), mcp-server-git (status, commit, checkout), mcp-server-terminal (run_command)}

State Persistence & Memory:
{Sliding Window Context with Ephemeral SQLite Checkpointing}