A modular, model-agnostic toolkit and collaboration framework designed to stop AI coding assistants from wandering into existential loops and breaking production codebases.
Why Coding Agents Need Strict Discipline
We have all been there: you ask an autonomous coding assistant to refactor a single CSS class, and ten minutes later it has migrated your database to MongoDB, rewritten your build pipeline, and invented its own logging format.
Agents Boilerplate (github.com/milkboyinchina/agents-boilerplate) provides standardized, model-agnostic guardrails and multi-agent protocols. Built for runners like Google Antigravity and OpenCode, it gives LLM pair programmers clear boundaries, surgical jump-points, and self-auditing habits.
What Is Inside the Toolkit
- Traffic-Light Teaming Protocol: Green Team plans with surgical anchors, Amber Team executes code within bounded scopes, and Red Team audits for regression slips.
- Concise Question Protocol: Formats clarifying queries into monotonic tokens (
Q1-a/b/c) for maximum token efficiency and rapid alignment. - Clean Session Handoffs: Checkpoint system that saves session context, open questions, and in-flight plans into lightweight local state files.
- Surgical Code Anchoring: Forces agents to cite precise symbols, approximate line numbers, and mutation types before changing single lines of code.
Quick Setup
# Clone the boilerplate framework into your project
git clone https://github.com/milkboyinchina/agents-boilerplate.git .agents-kit
# Run verification and initialize protocol workspaces
python3 .agents-kit/scripts/init_teams.py --checkFull documentation and specs available at GitHub: milkboyinchina/agents-boilerplate.
