CODEFORGE — Multi-Agent Code Generation System

The Pipeline

Six agents, one goal:

Working code. Each agent hands its output to the next, with feedback loops that fix problems before you ever see them.

STEP 01

ORCHES-
TRATOR

Reads the task → normalizes it

Detects programming language (defaults to Python), summarizes the task clearly for downstream agents, and initializes the shared state.

Groq LLM JSON output Language detection

STEP 02

PLANNER

Turns task into a blueprint

Creates a numbered step-by-step implementation plan covering required functions, classes, libraries, edge cases, and error handling.

5–10 steps Edge cases Library selection

STEP 03

CODER

Writes production-quality code

Uses the plan to write complete, runnable code. Can call Tavily web search to look up APIs and documentation. Reads feedback from Reviewer to fix previous attempts.

Tavily search Tool calling Feedback-aware

STEP 04

REVIEWER

Scores and approves — or sends back

Scores code 1–10 on correctness, completeness, quality, error handling, and security. Score ≥ 7 passes. Below that, returns to Coder with specific feedback. Max 3 rounds.

Threshold ≥ 7 Up to 3 loops 5 criteria

STEP 05

EXECUTOR

Runs the code, captures output

Writes code to a temp file and executes it via subprocess with a 30-second timeout. Non-Python languages (JS, Java, etc.) are marked successful with a note.

subprocess 30s timeout Isolated process

STEP 06

DEBUGGER

Fixes runtime errors automatically

If execution fails, receives the code + error traceback, diagnoses the root cause, and returns a complete fixed version. Loops back to Executor. Up to 3 debug attempts.

Root cause analysis Up to 3 attempts Auto-retry

Integrations

Three ways to run it

Terminal, visual canvas, or observability dashboard — pick your interface.

⚡

LangGraph Terminal

Run the full 6-agent pipeline from your terminal. Watch each agent work in real-time with color-coded output. LangSmith traces every step automatically.

            # run it

            python3 main.py

            # run tests

            python3 tests/test_suite.py

View on GitHub →

🔍

LangSmith Tracing

Every agent call, token count, latency, and LLM response is traced to LangSmith under project CODEFORGE. Full observability, zero config.

● CODEFORGE — Project

Runs: 127 | Avg latency: 4.2s

Tokens/run: ~3,200 avg

smith.langchain.com →

Open LangSmith →

🎨

Flowise Visual Canvas

Browser-based visual pipeline. ChatGroq + BufferMemory + CODEFORGE Agent connected with real edges. Chat with the agent directly in the Flowise UI.

CODEFORGE — Code Generation Agent

[ChatGroq] → [Memory] → [Agent]

Type a task → get code instantly

Open Canvas →

Live Demo

See it run

This is what CODEFORGE outputs when you ask it to implement a binary search tree.

CODEFORGE · python3 main.py

❯ python3 main.py

Task: Implement a binary search tree with insert, search, and delete

[ORCHESTRATOR] Task analyzed → Language: PYTHON

[PLANNER] Created 8-step implementation plan

[CODER] Generating code (iteration 1)...

[CODER] → searching: "Python BST delete node algorithm"

[CODER] Code generated (847 chars)

[REVIEWER] Score: 9/10 · Approved ✓

[EXECUTOR] Running Python code...

✓ Execution successful

─────────────────────────────────────

BST Tests: insert 5,3,7 → search(3)=True

delete(3) → search(3)=False ✓

─────────────────────────────────────

Total time: 8.4s · Agents: 5 · LangSmith: traced

❯

Quick Start

Up in 4 steps

Clone, configure, and run. No Docker, no server setup.

Step 01

Clone the repo

Pull the project from GitHub and navigate into it.

          git clone https://github.com/

            vignesh2027/CODEFORGE

          cd CODEFORGE

Step 02

Install dependencies

Python 3.10+ required. All deps in requirements.txt.

pip install -r requirements.txt

Step 03

Set your API keys

Copy the example env file and fill in your keys.

          cp .env.example .env

          # edit .env:

          GROQ_API_KEY=gsk_...

          TAVILY_API_KEY=tvly-...

Step 04

Run CODEFORGE

Start the agent or run the full test suite to verify everything works.

          # run the agent

          python3 main.py

          # run all tests

          python3 tests/test_suite.py

Capabilities

What makes it different

🔁

Autonomous review loops

Reviewer scores code 1–10. Sub-7 scores trigger automatic rewrites with targeted feedback — up to 3 rounds before forcing through.

🕷️

Real-time web search

Coder agent calls Tavily Search when it needs current documentation, API references, or package-specific implementation patterns.

⚡

Groq inference speed

Llama 3.3 70B on Groq hardware. Full pipeline runs complete in under 10 seconds for most tasks — 5–10× faster than OpenAI.

🔍

LangSmith observability

Every agent call is traced — token counts, latencies, prompts, responses. Debug exactly what each agent did and why.

🛡️

Isolated code execution

Code runs in a separate subprocess with a 30-second hard timeout. Errors are caught and passed to the Debugger, not the user.

🧩

Multi-language support

Detects Python, JavaScript, TypeScript, Java, Go, Rust, and more. Python gets full execution; other languages get complete reviewed code.

Code that writes
itself.

Six agents, one goal:

How the state flows

Three ways to run it

See it run

Up in 4 steps

What makes it different

Built with

Code that writesitself.

Six agents, one goal:

How the state flows

Three ways to run it

See it run

Up in 4 steps

What makes it different

Built with

Code that writes
itself.