← cd /projects
SWARM_AGENT.EXE

Swarm Agent

A local AI coding agent built on the ReAct+TDD architecture. Implements a Code → Verify → Reflect loop using Ollama, with long-term memory, tool execution, and a role-based Architect → Developer → Utility workflow.

Go Ollama AI Agent ReAct Local AI TDD
source code
README.md

A Self-Verifying Local AI Coding Agent

Swarm Agent implements a Code → Verify → Reflect loop — it never outputs code it hasn’t proven correct through execution. Built in Go, powered by Ollama (qwen2.5-coder:32b), and installable in 30 seconds.

Key Features

  • Verified Output — Executes and tests code before returning results
  • Self-Learning — Long-Term Memory (LTM) persists lessons across sessions
  • Tool Ecosystem — Python REPL, Performance Checker, FileTool
  • Role System — Architect → Developer → Utility with specialized model prompts per role
  • Zero-Config — Works out of the box, fully customizable via config

Architecture

Built on the ReAct (Reasoning + Acting) pattern extended with TDD verification:

  1. Think — Architect role plans the approach
  2. Code — Developer role writes the implementation
  3. Verify — Executes code through Python REPL or test runner
  4. Reflect — On failure, LTM records the lesson and retries

Quick Start

curl -sSL https://swarm.sh | sh
swarm "Write a Python function that calculates fibonacci numbers"

The agent thinks, codes, verifies, and returns working code.