Last year, 2025, the Y Combinator and Discord world I am in was raving about Claude Code. At that time, I was skeptical because my usage of LLMs in general had been mixed at best. Maybe Copilot integrated into IntelliJ or Visual Studio, but that was about it. When June rolled around, I installed Claude Code, which was fairly new at the time, and really dove in. I'm sure most folks go through this, but the initial phase of using Claude Code in the CLI and being able to iterate over code as fast as you can think of an idea is intoxicating. I wasn't under the illusion I could generate a 2MM ARR business overnight and not make a mistake; I was more interested in how much you could create and how fast it accelerated learning.
The Early Days
In the past, I would spend days or weeks reading documents and using trial and error to figure out best practices for an individual language; now I could brute force that learning, throw away code if it was bad and start over. No longer was I going to get stuck just choosing a language; now I could use Python and Flask or switch entirely to Rust if I was not feeling that path. I remember the frustration of going through docs, getting stuck on setup or not understanding the build process or not being able to debug because I didn't understand the ecosystem. That has all gone away.
The Progress and Multiple Agents
Then came competition. I have installed and worked with the following agents: Gemini CLI, Codex, Antigravity, Cursor, and a few others. I've delved into Minimax, GLM, QWEN, and various other open-source models. All cool, but I kept falling back to Claude Code. At the same time, I was adding plugins, agents, skills, MCP servers, and my workflow became bloated. I didn't know it at the time, but I needed to evolve how I used these tools to: 1. minimize cost; 2. be more effective.
Then near the end of 2025, Clawdbot (aka OpenClaw) came out and it was burning up the tech sphere. While I didn't install it right away, I did do some deep diving and wanted to learn about the internals, and that is when I stumbled upon the agent powering Clawdbot, Pi Agent: https://pi.dev.
That was the moment when it clicked for me: what a harness was, what its relationship to an LLM was, and how, when their powers are combined, we create an agent. The harness is the conductor utilizing an LLM, tools, and various other components to create the agent.
So, sell me on Pi
In January 2026, I'm tired of the restrictiveness of Claude Code and I have a more mature workflow with all the agents I have tried out; I want to fully customize how I interact with agents. So I load the Pi agent for the first time and let it go all out. No more asking me if I am sure; it just does it. With that great power came great responsibility, so I had to write or install what Pi calls an extension: My Tools. Once I blocked rm -Rf from the agent, I have not looked back. I am able to be nimble with the usage of my context, fully customize my workflow and how I use the agent, and really ask the question, do I need this MCP, extra connection, agent or skill?
Customizing, YOLO by default
If you need to connect to Confluence, ask Pi to build you an extension or skill for that. It's that simple; you can obviously look for already-built extensions or packages, like confluence https://www.npmjs.com/search?q=@pi-bot/extension, as the community is large, but you don't need to. What you really need to do is understand your workflow and what works best; don't let a coding harness and ultimately your agent define that for you. You know yourself best.
System prompt
The system prompt in Pi is very powerful, yet the token usage is minimal.
Token Usage Comparison (Snapshot)
| Metric | Pi.dev Agent | Claude Code |
|---|---|---|
| System Prompt Size | ~200–300 tokens | ~10,000+ tokens |
| Startup Overhead | ~0% of context | ~10% of context |
Model flexibility
This is the part that was missing for me with proprietary coding agents. With limits becoming more restrictive and models all converging on the same performance, I want to be able to swap out the LLM when I want. Today I could use ChatGPT 5.5 x-high for a work project and tonight switch to Gemini 3.1 Pro or Opus 4.6 for another task. The simple command /model gives me that flexibility without having to change my entire workflow.
Discord and the community
Ultimately, the health of any project is its community. On Discord and with OpenClaw, there are a lot of users of Pi, so much so that Mario and others are pretty swamped with requests and improvements. I have found the community to be responsive, helpful, and a great place to learn about models, agents, and coding.
The Ask
Try it. Download Pi https://pi.dev/download/cli and check out the Discord https://discord.com/invite/3cU7Bz4UPx. Let it rip on a project and let me know what you think. The more we understand about the tools we are using as software engineers, the better our ecosystem and ultimate product will be in the end.
-M