Agentry 🤖

Multi-Agent Code Review System

Write code, not tickets — AI agents collaborate to build, test, and improve your code through real CI/CD.

---

🌟 Overview

Agentry is an autonomous code generation and improvement system. You describe a task in natural language, and a team of AI agents — Proposer → Tester → Critic → Improver — iteratively writes, tests, criticizes, and improves code against real CI/CD metrics (not simulated scores).

Every LLM call is tracked: tokens, cost, latency, and quality score. The dashboard shows the full pipeline in real time.

How It Works

graph LR
    A[You describe a task] --> B[Proposer writes code]
    B --> C[Tester writes tests]
    C --> D[CI: tests, lint, pitest]
    D --> E[Critic evaluates metrics]
    E --> F{Passes quality gate?}
    F -->|No| G[Improver fixes code]
    G --> C
    F -->|Yes| H[✅ PR ready]

Loading

---

🏗️ Architecture

Agentry is a multi-module Gradle project following a feature-based modular structure:

Module	Responsibility	Dependencies
agentry-core	Domain entities, agent interfaces, pipeline logic	(pure Java)
agentry-persistence	JPA entities, Liquibase migrations, PostgreSQL	core
agentry-api	REST controllers, DTOs, API layer	core, persistence
agentry-ci-gateway	GitHub Actions webhooks (POST /ci-callback)	core, persistence
agentry-cli	Spring Shell console interface	core
agentry-dashboard	React frontend host (future)	api
agentry-app	Spring Boot entry point, configuration	all modules

Tech Stack

Component	Technology	Version
Language	Java (OpenJDK)	21 LTS
Framework	Spring Boot	3.4.x
Build	Gradle (Kotlin DSL)	8.12
Database	PostgreSQL	16
Migrations	Liquibase	4.29
CI/CD	GitHub Actions	—
LLM	Anthropic Claude API	latest
Testing	JUnit 5 + Mockito + Testcontainers	latest

---

🚀 Getting Started

Prerequisites

• Java 21 (auto-downloaded by Gradle via foojay toolchain resolver)
• PostgreSQL 16+ (for persistence layer)

Build & Test

# Clone the repository
git clone https://github.com/andriibats/agentry.git
cd agentry

# Build the project (Gradle auto-downloads JDK 21)
./gradlew build

# Run all tests
./gradlew test

# Run a single module's tests
./gradlew :agentry-core:test

# Start the application
./gradlew :agentry-app:bootRun

Quick Start (CLI)

# Run the CLI without starting the full web app
./gradlew :agentry-cli:bootRun

---

🧠 Agent Roles

Role	Prompt	Output
Proposer	"Write minimal working code"	Implementation files
Tester	"Write tests — be critical, find edge cases"	Unit + mutation tests
Critic	"Evaluate CI metrics"	Score + feedback
Improver	"Improve code based on feedback"	Fixed implementation

Task Budget

Each task has a token budget controlled by the user. Agents autonomously decide how to spend it — more on the first draft vs. multiple refinement rounds — making trade-offs between quality and cost.

"Remaining budget: 1,850 tokens. Account for this when deciding whether to continue improving."

---

🗺️ Project Roadmap

Stage	What	Status
Scaffolding	Project setup: multi-module Gradle, CI/CD, Claude Code AI infra	✅ Done
Stage 1	Core orchestration + agents + logging	🚧 Active
Stage 2	Git integration + real CI/CD	📋 Planned
Stage 3	Dashboard (React + SSE)	📋 Planned
Stage 4	Automation & hooks	📋 Planned
Stage 5	Multi-provider agents	📋 Planned
Stage 6	Memory autodream	📋 Planned
Stage 7	Real project integration	📋 Planned
Stage 8	Authorization & BYOK	📋 Planned
Stage 9	LLM response caching	📋 Planned

---

🛠️ Development

Project Structure

agentry/
├── agentry-core/             # Domain logic (pure Java)
├── agentry-persistence/      # JPA + Liquibase + Postgres
├── agentry-api/              # REST API
├── agentry-ci-gateway/       # GitHub webhooks
├── agentry-cli/              # Console interface
├── agentry-dashboard/        # Frontend (future)
├── agentry-app/              # Spring Boot app
├── .claude/                  # Claude Code configuration
│   ├── CLAUDE.md             # AI context
│   ├── agents/               # AI agent definitions
│   ├── skills/               # Custom skills
│   └── memory/               # Persistent project memory
├── docs/                     # Documentation
│   ├── specifications/       # System design documents
│   └── plans/                # Implementation plans
└── .github/workflows/        # CI/CD pipelines

Code Conventions

• Java 21 features: records for DTOs, text blocks, pattern matching
• No Lombok: explicit constructors and getters (or records)
• Testing: JUnit 5 + Mockito + Testcontainers
• Database: ddl-auto: validate — all changes through Liquibase

---

🤝 Contributing

This is an open-source pet project. Ideas, issues, and PRs are welcome!

• Spec & design documents live in docs/specifications/
• Implementation plans live in docs/plans/
• AI-assisted development is encouraged via .claude/ config

---

📄 License

MIT

---

_{Built with ❤️ by @andriibats}