coding-agents-loop

Multi-agent coding system with 3 coordinated agents (Orchestrator, Coder, Reviewer) running in a review loop until the code is approved.

┌─────────────────────────────────────────────────────────────┐
│                    docker-compose                           │
│                                                             │
│  ┌──────────────┐    ┌──────────────┐    ┌──────────────┐  │
│  │  CLIProxyAPI │    │   LiteLLM    │    │    Agents    │  │
│  │  :8080       │◄───│   :4000      │◄───│  (Agno +    │  │
│  │              │    │              │    │  Claude CLI) │  │
│  │ Gemini OAuth │    │ orchestrator │    │              │  │
│  │ Claude OAuth │    │ → Gemini 3.1 │    │ Orchestrator │  │
│  │ Codex OAuth  │    │ coder        │    │ Coder        │  │
│  │ (multi-acct) │    │ → GPT-5.3    │    │ Reviewer     │  │
│  └──────────────┘    │ reviewer     │    └──────────────┘  │
│                      │ → Sonnet 4.5 │                       │
│                      └──────────────┘                       │
└─────────────────────────────────────────────────────────────┘
              │
              ▼ /workspace (volume condiviso con host WSL2)

Stack

Component	Role
CLIProxyAPI	Go proxy — unifies Google Cloud Pro (Gemini), ChatGPT Pro (Codex), Claude via OAuth. No API keys needed.
LiteLLM	OpenAI-compatible model router mapping aliases → backends via CLIProxyAPI
Agno Framework	Python framework coordinating the 3 agents as a Team with mode="coordinate"
Claude Code CLI	Execution engine for all 3 agents

Agents

Agent	Model	LiteLLM alias	Role
Orchestrator	Gemini 3.1 Pro Preview	orchestrator	Decomposes tasks, assigns to Coder, iterates until approved
Coder	GPT-5.3 Codex	coder	Implements code using filesystem/shell/git tools and MCP servers
Reviewer	Claude Sonnet 4.5	reviewer	Reviews output, returns issues in Italian or approves with "APPROVATO"

The Loop

User Task → Orchestrator → Coder → Reviewer
                ↑                      |
                |  APPROVATO           | issues found
                └──────────────────────┘
                     ↓ (loop, max MAX_REVIEW_ITERATIONS, default 5)
                  Coder (fixes issues with context)

Language rule: agents communicate with each other in English (more efficient for LLMs). All final output to the user is in Italian.

Prerequisites

• Docker Desktop with WSL2 backend
• tmux: sudo apt install tmux
• Active subscriptions: Google Cloud Pro, ChatGPT Pro, Claude Pro

Setup

git clone https://github.com/Thomas-Piva/coding-agents-loop
cd coding-agents-loop

cp .env.example .env
cp cliproxy/config.example.yaml cliproxy/config.yaml

# Authenticate each provider (run once)
make login-gemini
make login-claude
make login-codex

# Start all services
make up
make ps

Usage with Tmux

chmod +x tmux-start.sh

# Open monitoring session (no task)
make tmux

# Run a task
./tmux-start.sh "Build a Flask REST API with /health and /items endpoints"

# Use a template
./tmux-start.sh "template:webapp"
./tmux-start.sh "template:api_service"
./tmux-start.sh "template:refactor"

Tmux navigation: Ctrl+B arrows to move between panes, Ctrl+B Z to zoom, Ctrl+B D to detach, tmux attach -t coding-agents to reattach.

Available Templates

Template	Description
template:webapp	Full-stack web application (frontend + backend + DB)
template:api_service	Production REST API with auth, validation, OpenAPI docs
template:refactor	Analyze and refactor existing code in /workspace

Extending the System

• New agent: create agents/agents/myagent.py, add to dev_team.py, add alias in litellm/config.yaml
• New tool: create agents/tools/mytool.py, import in the relevant agent
• New MCP server: add entry in agents/mcp_servers/config.json
• New template: create agents/tasks/mytemplate.yaml with fields: name, description, orchestrator_prompt, coder_focus, reviewer_checklist

Makefile targets

make up            # Start all containers
make down          # Stop containers
make build         # Rebuild agents image
make tmux          # Open tmux monitoring session
make run TASK="…"  # Run a task
make logs          # Tail all logs
make ps            # Show container status
make clean         # Clear workspace
make login-gemini  # OAuth login for Gemini
make login-claude  # OAuth login for Claude
make login-codex   # OAuth login for Codex