Agentic Engineering

A working demonstration of AI agents as a software development team — autonomously implementing features, running tests, opening pull requests, and reviewing code, while a human architect defines the vision and reviews the output.

---

The Philosophy

The role of a software engineer is shifting. The most valuable skill is no longer writing code — it's thinking clearly about what needs to be built and why.

AI agents can implement, test, review, and iterate. What they can't do is decide what matters, understand business context, or make architectural trade-offs. That's the human's job.

Engineers think. Agents execute. Humans review.

The engineer's value is in designing systems, defining quality standards, and building the infrastructure that makes agents reliable. Agentic workflows are where the industry is heading — this repo is my exploration of what that looks like in practice.

---

The Pipeline

Issue Agent scans the codebase for unimplemented features
    → opens a GitHub issue for each one
    ↓
Webhook server receives the issue event
    ↓
Orchestrator coordinates the agents
    ↓
Coding Agent (Claude SDK)
    → reads the codebase
    → implements the feature
    → runs integration tests
    → commits and pushes to a branch
    → opens a pull request
    ↓
Review Agent (GitHub App bot)
    → reads the diff
    → evaluates against defined rules
    → approves or requests changes
    ↓
If changes requested:
    → Coding Agent reads feedback and fixes
    → Review Agent reviews again (max 2 loops)
    → If still failing: labels PR "needs-human-review"
    ↓
Developer reviews the PR and merges

The developer's only manual step is reviewing and merging the PR.

---

Agent Architecture

Agent	Type	Purpose
Issue Creator	Deterministic script	Scans codebase for unimplemented features, opens GitHub issues
Coding Agent	Claude SDK agentic loop	Implements features end-to-end — reads code, writes code, runs tests, commits
Review Agent	Claude SDK agentic loop	Reviews PRs against defined rules, approves or requests changes
Orchestrator	Coordinator	Runs the implementation → review → fix loop, enforces max iterations
Webhook Server	Hono HTTP server	Receives GitHub events, triggers the orchestrator

Tools available to agents

• read_file — read any file in the codebase
• write_file — write or update files
• run_bash — execute shell commands (build, test, etc.)
• git_create_branch, git_commit, git_push — git operations
• submit_review — post a PR review via GitHub API

---

Tech Stack

Layer	Technology
AI Agents	Claude Agent SDK (@anthropic-ai/sdk)
Webhook Server	Hono + @hono/node-server
GitHub Integration	@octokit/rest
Local Tunnel	ngrok
Runtime	Node.js + TypeScript (tsx)
Underlying Program	Solana / Anchor
CLI	Rust (clap, anchor-client)

---

The Underlying Project

The agents are building a Solana token program — a custom on-chain token implementation with mint, transfer, burn, and balance functionality. The program is written in Rust using the Anchor framework.

The Solana project exists to give the agents a real, non-trivial codebase to work in — with IDL-driven code generation, integration tests against a local validator, and strict conventions. It's a meaningful test of whether agents can maintain quality in a complex environment.

---

Getting Started

Prerequisites

• Node.js 18+
• Rust + Anchor CLI (for the Solana program)
• ngrok account (for local webhook testing)
• Anthropic API key
• GitHub Personal Access Token

Setup

cd agent
npm install
cp .env.example .env
# fill in your credentials

See agent/README.md for full setup instructions including ngrok and GitHub webhook configuration.

Running the pipeline

Start the webhook server:

cd agent && npm run webhook

Expose it via ngrok:

ngrok http 3000

Trigger the full pipeline:

cd agent && npm run issues   # opens GitHub issues for unimplemented commands

Then watch the agents work.

---