The AI Magpie

How I've mostly avoided becoming one.

For ~6 months I’ve managed to avoid the temptation to adopt the shiny things. In today’s AI world, that can be quite tricky. Before I share how I’ve avoided the temptation, we should explain what an AI Magpie is:

AI magpie
/ˌeɪˈaɪ ˈmæɡpaɪ/
noun

Definition:
A person who collects shiny AI tools, models, agents, workflows, prompts, dashboards, and half-finished architectures with the sincere belief that each one may eventually become “strategically useful.”

Usage:
“I was supposed to be finishing the roadmap, but then I discovered another local inference framework and entered full AI magpie mode.”

I’ve mostly managed to dodge this trap. Since late December, I’ve been working with the same software factory. Built mostly from scratch by me, taking heavy inspiration from Mario Zechner’s early posts on Pi and a small handful of basic, but powerful patterns to ensure quality (multi-judge evaluations, deterministic pipelines etc.).

What I built wasn’t intended as a personal assistant. It was an agent with one goal:

Enable a non-technical product manager to deliver software that isn’t slop.

It was simple and it worked. I called it Arnor.

And so I’ve used that same ‘software factory’ for over 5 months. It has served me well. It’s helped me build multiple tools for myself and built the entire backend for ColdAnvil.com.

The lure of shiny tools

Over those 5 months, it’s not been easy. OpenClaw got even better. Then everyone and their dog decided to build their own Claw. Jensen Huang grew his own Claws. Cursor, Claude Code, Codex, Factory - all adding agentic features.

The list could go on. I follow the opensource AI community closely and there are wonderful things being built every day.

ColdAnvil.com and my other projects have kept me focused and away from the shiny allure of new AI software. But… as the months have moved on, I’ve felt my home-brewed agent age. It has some automation, but I spend many hours a week glued to the screen guiding, questioning, challenging my agent. Not just on product work, but actively challenging it’s engineering. If I’m to hit my goal of launching three products in three months, I need to spend less time babysitting Arnor.

Enter, Hermes. My shiny nemesis.

If you aren’t aware of Hermes, it’s often been touted as the OpenClaw killer. Beyond the hype, it is a legitimately strong agent. It iterates itself based on your needs, includes agent/sub-agent workflows and it’s got all the tools and skills baked in.

It’s opinionated over what an agent is, but still manages to meet a broad spectrum of needs.

This left me, for the first time in months, tempted to move away from my home-brewed agent. Arnor was a bottleneck and I needed something faster.

Thankfully, the enduring lessons of product management and first principles thinking saved me from myself.

One human skill to rule them all.

The product skill that helps most in these situations is product judgment: the ability to separate interesting possibilities from meaningful opportunities. In the age of AI, that judgment matters more than ever, because the solution space has exploded. There is always another model to test, another agent to try, another workflow to automate. Product discipline and judgement brings you back to first principles: what problem are we solving, for whom, why does it matter, and what would count as real progress?

So, as a result, I asked myself ‘What do I actually need?’

There are many agents, but this one is mine.

The answer was incredibly simple. I need Arnor V2. I need the agent I alread had, but faster and more autonomous. All whilst achieving the same, if not better output.

And so that’s exactly how I’ve spent the last week. I paused on ColdAnvil.com, to rewrite Arnor entirely.

It’s still based on Pi, but thanks to the flexibility of Pi, it’s grown from a simple agent, to become my own AI software studio:

One human-facing orchestrator backed by specialist agents for product, engineering, architecture, design, data, review, and verification.

Instead of a generic coding agent, it runs product work through explicit pipelines: gather context, plan, decompose, assign scoped agents, implement in parallel. Improved deterministic checks, perform independent technical and product verification, then ship a PR or escalate for the feedback loop.

It still doesn’t have the bells and whistles of Hermes, but it’s perfect for my workflow and my projects.

Did you notice ‘mostly avoided’?

This is the part where I get honest. I have avoided constant tool switching, but I can’t say I’ve avoided shiny new toys.

Two key reasons:

1. I’m running this on very shiny local infrastructure with 300gb VRAM and NVIDIA Blackwell GPUs, escalating to frontier models where needed.

2. There are a number of wonderful open source projects which I’ve selected over time based on my goals and their ability to fill a gap in my stack

So take my preaching with a large pinch of salt. But if you’re going to do one thing, make sure you remind yourself of what you actually want to achieve. It might stop you from trying every tool and maybe help you focus on the few that help you succeed.