跳轉至

🤖 Autonomous Agents

Intelligence Maximization Component "You hire a team, not just a chatbot."

Boring-Gemini operates as a Multi-Agent System (MAS). When you give a complex task, it's not just one LLM trying to do everything. It creates specialized personas to handle different aspects of the work.

👥 The Squad

1. Orchestrator (The Boss)

  • Role: Project Manager.
  • Job: Understands your goal, breaks it down into steps (task.md), delegates to other agents, and ensures the final result matches the requirement.
  • Tool: boring_multi_agent

2. Architect (The Planner)

  • Role: Senior Engineer.
  • Job: Designs the solution, creates implementation_plan.md, checks for architectural consistency, and reviews major changes.
  • Tool: boring_prompt_plan

3. Coder (The Builder)

  • Role: Software Engineer.
  • Job: Writes code, fixes bugs, and runs tests. Focuses on execution.
  • Powered by: Vibe Coder capabilities.

4. Reviewer (The QA)

  • Role: Quality Assurance.
  • Job: Reviews code for bugs, security issues (security_scan), and performance problems (perf_tips).

🧠 System 1 vs. System 2 Reasoning

Boring-Gemini uses a Dual-Process thinking model to optimize cost and performance:

  • System 1 (Fast): Used for routing tools and simple edits. It relies on the model's intuitive "next-token" prediction.
  • System 2 (Slow): Automatically triggered for complex tasks (architecture changes, big refactors). It uses ReasoningState to perform internal deliberation before calling any tools.

🔄 The Autonomous Loop

Unlike standard chatbots that stop after one reply, Boring runs in a Loop:

  1. Plan: Architect creates a plan.
  2. Execute: Coder writes the code.
  3. Verify: Reviewer checks the code and runs tests.
  4. Fix: If tests fail, Coder fixes them (Loop continues).
  5. Done: Only stops when criteria are met.

🚀 Vibe Coder Usage

Use sequentialthinking explicitly, or let the boring-route decide.

# Complex task -> Triggers Planner/Orchestrator
boring-route "Help me design and build a new user auth system"
# 🎯 Routes to boring_prompt_plan (Planning)

# During the process, agents might call:
# - sequentialthinking (to solve logic puzzles)
# - context7 (to check docs)

🛠️ Configuration

Enable Multi-Agent features in .boring.toml (Standard/Full profile recommended).