echo "=== common SDK locations ==="; for d in "$HOME/Android/Sdk" "$HOME/android-sdk" "/opt/android-sdk" "$HOME/.android" "$HOME/Documents/android-studio" "$HOME/Android"; do echo "--- $d ---"; ls "$d" 2>&1 | head; done; echo; echo "=== ANDROID_HOME / SDK_ROOT env ==="; echo "ANDROID_HOME=$ANDROID_HOME"; echo "ANDROID_SDK_ROOT=$ANDROID_SDK_ROOT"; echo; echo "=== sdkmanager search ==="; find /home /opt -name 'sdkmanager' 2>/dev/null; echo; echo "=== adb ==="; find /home /opt -name 'adb' -type f 2>/dev/null; echo; echo "=== platforms (android-35) ==="; find /home /opt -type d -name 'android-35' 2>/dev/null
My Pi setup (August 13th 2026)
During work today we had a discussion around coding agents and harnesses. At Volvo, most of us use Github-copilot, but a few people use lang-chain and Phoenix AI observability tools.
A lot of people aren’t really familiar with the rapid pace of the AI development landscape. I have two goals for this article.
-
Give a quick summary of the unintuitive parts of AI agent development
-
Show off my Pi setup, and encourage people to move to open-source tools.
Something changed in Winter 2026
We’ve heard about 10x developers and AI making engineers obsolete for over 4 years. People really thought GPT-2 was world ending, once upon a time. There’s a lot of smoke and mirrors - and I just want you to throw away the hype for a seconds.
Just internalize this: Earlier this year coding agents popped off.
Suddenly, we can have pair programming assistants who can autonomously look at functions and fix them, structure programs, accomplish TODOs, and fix tests. This is really cool.
Again, ignore the AI hype train. We finally got a taste of what we were promised in 2023… Finally in 2026.
Open source is king
Open source has been on everyone’s lips.
This year alone, it has been proven dozens of times that open source Chinese models can outcompete (on both price and quality) American-made models. I’m not here to talk about the implications, or reasons of this. I just want to bring it up.
The downside is privacy. But there’s some alternatives…
Openrouter.ai links up users with companies who host open source models (for a slight upcharge). There’s a couple of companies with strict Zero-Data-Retention (ZDR) policies.
You can also get a local model, and use that instead. You can either use a small one, or a big one that’s been quantized down.
Let’s get into my PI
_How do I verify _
_Why a smaller, weaker harness? Why not something like Claude that has MCP or Codex with a million features? _
One of the irritating things about utilizing LLMs in work is that you actually have surprisingly little control of a lot of the features. At a baseline, fundamental level - you ask an AI twice, you get two different answers. Most of the time - this is fine. 5 different runs will have a lot of overlap, especially for well-defined tasks. However, in order to increase reliability, it’s extremely important to reduce as much random fluff as possible, especially when it’s not relevant to your work. It’s also necessary to customize your tools.
Why don’t you use MCP? What is the benefit of Bash?
A technical reason is that a lot of these harnesses use bespoke features that are custom fine-tuned per model+harness. For example, if you ask "claude what does this document say" - it has its own internal MCP connector.
However, Pi (and my linux workspace) uses bash.
The benefit of this is obvious when we think from first principles.
-
LLMs need lots of training data for EVERYTHING they do. They are not logical, they have no brain, they just follow patterns.
-
Bash was created in 1989. A lot of the shell commands, syntax and logic have their roots in 1983 (the start of Gnu) and even earlier (older text-only terminals). We have decades of high, quality data and tutorials.
-
Claude MCPs are new. Under 4 years old. The data is not public, so they need to rely on private data.
-
This can be really good for Anthropic models, like Claude. But Deepseek, and other Chinese open-source models doesn’t know how to use them, so it’s worse.
-
Ergo: Bash probably is a bit better for regular agents.
But it gets even more interesting when we think about what happens for multi-step work.
-
You can chain bash commands together using | pipe commands. It’s trivial to do things like go over 100 folders, look at every file in a subfolder, delete the file if the first 10 characters are a specific string, then return the result.
-
I haven’t used Claude Opus 5 for this; But I would expect it wouldn’t do as well, since the underlying connector is something bespoke. If the situation isn’t in its training data, I don’t know if it could logically find out how to do it without thinking.
BUT: What if your use case is very specific? What if you need to have a custom script? It’s trivial to add whatever bash command you want, then call it. Customizing Claude MCP is much harder.
Recommended reading: Smaller harnesses