Hello!
Let’s see if this thing works.
Testing… tetsing.. it does!
Wow, been a long time since I actually wrote something myself that wasn’t, you know, regurgitated by a clanker. (yes, I get to call it a clanker now, you can too if you accept the message of the Pi.dev intern into your heart)
Why am I writing this thing again, anyways? My attention span is completely messed up from all of this agent stuff… and it just occurred to me that I might make a mistake and make myself look illiterate if I’m not careful, so I better do my best job here to focus and push through with this, or else…
Yes, yes, now I remember. I wrote a Linkedin/X post recently where I shared some attention seeking content about how I’m now cool and all of that since I cancelled my Claude Max(20x) subscription and moved my personal workflow into Pi and presented (ta da!) the extension that came with that and which represents my workflow, now in Pi, called Pi-Brain. Mind you, I did have a Brain before all of this as well, it was mostly Claude specific stuff though, which is one of the reasons why I’m like this today.
This specialness resonated and one of the very famous, one of the best, if not the best, a good guy, AI influencers on Linkedin asked to share a bit more on why I did this, what I gained and what I lost, etc. Asim is his name, now I remember.
So, here we are. I am writing this now and I get to explain stuff.
Fighting the urge to just port thoughts into any agent, just paste them here and let you deal with the slop. But, I’ll try to be a good human in the loop and share slop that’s in my head instead.
Why Pi
Recently, I organized a conference called RubyConf Austria and on that conference one of the things we did, and there were many special moments, was host a panel discussion titled AI, Agents & the Future of Ruby: Chad Fowler, José Valim, Obie Fernandez, Armin Ronacher. On this panel we had some amazing people that I invited, and was thrilled they accepted, because I wanted to be close to their thinking and have some of their magic rub off on me, ask some questions that are on my mind etc.
One of the people that joined was Armin Ronacher, who is one of the thought leaders in the world of Agents and one of the three avengers in the Austrian school of AI or the Vienna school of Agentic Coding, which I’m trying to be a pupil of. I wanted to have Mario and Peter there as well, but Peter was busy with his move to SF and Mario just didn’t respond to my e-mail.
If you didn’t know, Pi is built and maintained by Mario and Armin, and it sits at the heart of OpenClaw which is built by Peter. Lately, it sits at the heart of many products building on top of open source, since Pi is available under the MIT license.
I still have written on my whiteboard “build close to what the model is being rewarded with” as a note from my conversation with Armin at the event. I don’t know why I wrote this as part of this post, I just read it on the board and now it’s here.
Anyways, I came to the event with some of my own thinking around building intent stores for the new SDLC (hence the brain), trying to achieve what Chad mentions around Phoenix Architecture in a way I understood it at the time, and some of the conversations I had around this topic, listening to the panel and in general following these people on social media for the past 7-8 months has shaped my thinking on the topic and provided new perspectives.
I was also very aware that my workflow is very Claude focused and dependent on Anthropic, which I did not like at all and I wanted to change, but I didn’t have the energy to do that at the time, had other obligations that I needed taking care of, needed to rest a bit after organizing the conference etc, so I kind of held it off for a while.
In the past 7-8 months, each time a new model from Anthropic came out, I had to adjust my workflow, change how I work with the model, some prompts stopped working as before and I was constantly making mini adjustments and my workflow was never really my workflow, I was using Claude’s harness and model as a typical consumer without much flexibility to make it my own, trying stuff with the brain was tiring at times and was constantly bringing Claude back to work how I intend it to work.
With Pi I am able to build a harness for myself and around my workflow how I like it to be built. I can replicate that for all of my projects and it will not change wildly with each rollout. That’s kinda the point of having a minimalistic harness that extends itself to fit you. It’s fun and dangerous and I get to feel like I’m a builder and not only a consumer.
Also, once I build it, I can use any model I want with it. So, the model adjusts to how I want to run it, not how Anthropic wants to or OpenAI or any other company.
Tomorrow, if I’m able to host an open source model (notice open source as a repeating theme), I can just continue using the harness I built for myself with Pi.
Why Kimi K3
Open source.
I want to see (good) open source models that I can host locally manifest into reality.
I choose to support companies that are close to this thinking and that support open source development of AI and agents.
Today that’s Kimi K3, tomorrow it might be something else from OpenRouter.
What I lost
The cushy feeling of being taken care of by Claude, having the entire ecosystem I can benefit from as a consumer.
Now I have to build my own mobile app with which I can access my sessions remotely, for example.
If I want to use ui.sh for design, I need to teach my Pi agent to use it. Everything is a conscious decision and I need to be careful and understand what I need and how I need it to work, then I need to build it myself (under my supervision and tailored to my needs).
Disclaimer
I still have a Claude subscription at work that I use.
I still work with OpenAI to develop AI features for the products I work on.
This is still true for some projects.
How did I handle my data once I cancelled my subscription
I didn’t. I’m lazy to worry about it at this very moment, might change once I get more energy to process stuff.
Maybe I export my data at one point.
Maybe it’s left where it is…