Actually, it’s not mine, it’s Mario Zechner’s. I’m talking about the wonderful pi coding agent. I’ve been using it for a month now – an eternity, considering the revolution that’s engulfing software development tools as of early 2026. pi is fantastic and I want to tell you why.
1) YOLO is hardcoded: no permissions are asked. Agents work at full speed, unless they are done or they have a question for you. Letting go of the permissions is difficult and disturbing, but it works. For me, it represented my entry into full agentic coding. It’s like going from micromanager to manager. You can still watch the agent work and catch things that are wrong and stop them (you stop it with Esc, not with ctrl+c). It is wise to run the agent inside a container in case it goes rogue and attempts to read your home directory or rm -rf you into oblivion. In any case: YOLO is the way to agentic speed. For good or else, the brakes are off.
2) Minimalism: only four tools: read file, edit file, write file, run a command. These four cover pretty much everything you need. If you need something else, ask the agent to write it for you. This means that your context window is not consumed by tools; the small tool surface also seems to make the agents smarter and rely on their knowledge of unix to get stuff done. I have no proof of this, but think of the best devs you know: they’re not using highly specific, vendor provided tools: they’re using tools that have been polished over time into a minimal and powerful core.
3) See-through: the client hides nothing from you. You can see every tool use. There’s no magic, no black box. It’s liberating to see what the agent is actually doing. You can also see it go astray and write a permanent instruction to avoid that pitfall, something that you cannot do when you don’t know what’s going on.
4) Elegant: pi is almost redis-like in how beautiful it feels when you use it. It’s been a very long time since I felt at home in a new piece of software.
Here’s what I use in pi every day:
- Ctrl+P to switch providers on the fly. Useful when you hit Anthropic rate limits or when gpt5.3 gets lazy and you need to resort to 5.2. It makes a lot of sense to configure what are your favorite models, so you can switch quickly.
- /compact so you decide when to compact the conversation.
- I avidly read the usage gauges at the bottom.
As for flow, while the line editing in pi is very well done, I still prefer to use my editor on the side to write most prompts, then copy and paste them into pi. This allows me to have 1) an even better experience of editing text; 2) a record of what I’ve been telling the agent, particularly if I have multiple agents going.
I know that the official agents of Anthropic and OpenAI have cutting edge features, but for the time being I’m not tempted to switch. Having a single, elegantly built agent to run any model is something very hard to give up.