It seems I need very little from agents

For the second month now, I have been living full-time in pi.dev as my main agent. Out of the box it has almost nothing except the loop, skills support, compacting, and four tools: read, write, edit, bash. Its basic idea is that any wish can be installed or implemented as an extension. To understand the scale of this extensibility: OpenClaw is built on top of it, and you can run Doom as a tool directly inside the TUI. That is nonsense, of course, but quite illustrative.

The funny thing is that I barely touched it after installation. It is nice to know that I can always extend it, but for most of the time I have been using it almost exactly as it launches the first time.

Turns out I do not need subagents, planning modes, teams, MCP, buddies, dreams, or all the other junk that comes out of the box in something like Claude Code or even OpenCode. For the first time I got close to the weekly limit of the $100 OpenAI subscription in this setup. I very rarely connect web search, and that is it. Everything else is on vibes. Like in that meme where all the complex concepts are in the middle of the curve, and the most straightforward approach is on both ends.

The author himself, Mario, amazes me with what he says. I highly recommend this episode of the Pragmatic Engineer podcast (probably my favorite podcast on the internet right now). It has a huge amount of great thoughts. In particular, I really like the idea of thinking in the Linux-way paradigm: creating minimal working tools instead of trying to solve all the problems of humanity and business with one product. And the thought of not adding fun nonsense to your products just because it can be done quickly is great too. He says roughly the same thing about Pi. And he also says that he uses it almost always in a minimal configuration, so I am not the only one like this.

The first thing Pi sells you on is the moment you see how fast the GPT model replies there. At that exact moment you understand how polluted other agents are with system prompts full of raccoons and goblins, plus bells and whistles that get injected into every request. The second thing is that it just works. So well that it is hard for me to nitpick. I did not write any hacks for it like I did for Claude Code, and I have not seen any visual bugs. Most importantly, Pi does not impose anything on you that may be unnecessary for the task. It gives you the simplest base you can trust. And that alone is priceless in a world where Claude Code is still a very popular agent.