How I Use Agents

Introduction

I've been coding for over 10 years now, and over the past few years coding has evolved rapidly. When you were stuck, you used to go to Stack Overflow, then to ChatGPT, and now to many agents.

My Setup

I wanted to share the current setup I use when contributing to open source and working on personal projects. I've tried hosted services like Replit Agent, agentic IDEs like Cursor and Antigravity, and CLI tools like Claude Code and Pi.

After all that, there's one clear winner: a little-known open-source CLI called OMP (oh-my-pi). OMP self-describes as "Pi with the batteries included," and that really sums it up.

I use it with the Codex Plus subscription because I like working with OpenAI's models the best right now, and it provides generous usage, although I also use local models occasionally. I used to use it with Claude and Gemini, and they work great too.

Contributing to Open Source

Contributing to open source is a great way to get experience and help with projects you love. When I want to contribute to open source, here's my process.

I start with an issue. Either one I have, or one from GitHub. I then ask ChatGPT Deep Research (although you can use any research AI, like Perplexity or Claude) to "Explain the issue, and come up with a solution if possible." After that, I follow up by asking it to generate a prompt for my AI agent, OMP. I paste it in and get to work.

I push to my clone and have ChatGPT, in the thread where it did Deep Research, check it. Based on its critiques, I re-prompt the agent until it's good to go according to ChatGPT. I then use OMP to clean up the git history and force-push, and I review the code manually (PLEASE DO THIS IF YOU ARE CONTRIBUTING TO OPEN SOURCE).

I then submit a PR, making sure I'm following the guidelines of the repo and disclosing my AI usage. This setup lets me be very productive.

Note: there are downsides to OMP. It does use a lot of tokens, and it requires a lot of updates, which is also a good thing.

Personal Agents

I also use agents for personal stuff. Email summarization. Daily briefings. Sending files from my MacBook remotely.

I've tried OpenClaw, NanoClaw, and Hermes, and Hermes was the best, but I quickly switched to something different. I use the Codex app with computer use now. OpenAI recently brought Codex to the mobile app, so I can now chat with my Mac remotely.

It also sends me daily morning briefings and can help with small coding tasks when I don't want to go through the hassle of OMP for something small, or I need to make changes from my phone. It's been an interesting competitor to the open-source command-line-based agents, but I think it's a winner.

Sadly, Claude and Gemini don't have an equivalent yet (Claude is close, Gemini lacks one entirely).