Pi is a really nice agent - Resolve.

I have always built my own working environment. I use nvim, customized through plugins. I use ergodox keyboards with a keymap that probably makes sense only to me. Oh, and I use Arch, by the way. Not because everyone should manage their computer this way, but because I want my tools to adapt to how I work—not the other way around.<br>There is an entirely reasonable alternative: let somebody else assemble the system, maintain it, and decide how its parts fit together. That removes responsibility and lets you concentrate on the work. The tradeoff is that you can only work within boundaries designed for the whole market. An open ecosystem lets you draw those boundaries yourself.<br>For years, I have assembled my development environment from tools people shared on the internet. That ecosystem has given me more than software. I have learned from the code, documentation, and ideas that other people published freely.<br>For about a year, Claude Code was an exception. I had used Anthropic’s coding harness since its early beta, its models were among the best for programming, and the whole thing did its job well. I had no reason to look elsewhere.<br>Then, in early 2026, Anthropic restricted Claude subscription credentials to its own products, preventing third-party harnesses from using them. API access remained available, but was metered separately.<br>The change clarified what Claude Max was: not a general model subscription, but a subscription to Anthropic’s products. That did not suit how I wanted to work. I wanted a harness I could shape myself, so I started looking for an open-source alternative.<br>It extends itself<br>I found what I wanted in the pi coding agent. Pi starts with a deliberately small core: read, write, edit, and bash. Everything else—extensions, skills, prompt templates, themes, and packages—is pluggable. It feels more like the open ecosystem I came from than a product trying to anticipate every workflow. Pi even encourages you to ask the agent to build extensions for itself, and ships comprehensive documentation and examples to support it. The tool is designed to help you build the tool.<br>Having used coding agents since their early days, I already had a good idea of where they struggled. The largest gap was exploration: building a global understanding of a codebase instead of accumulating a fragmented collection of files and snippets. I followed pi’s philosophy and resisted recreating every feature I had left behind. I would add only small, sharp tools for problems I had actually seen.<br>Tool descriptions are instructions: they tell the model what it can call, when to call it, and how to structure the call. Adding more tools therefore adds more instructions and more competing choices. In ManyIFEval, models became steadily less reliable as they were asked to follow more simultaneous instructions. LongFuncEval found the same pattern in function calling. There is no universal cliff or safe number. The practical rule is simply to keep the active surface as small as the work allows, and make every tool earn its place.<br>What I've built<br>Over the past six months, I've converged on this minimal set of extensions:<br>fork — lets one Pi session hand a focused task to another. Each sub-agent starts with a fresh context in its own tmux window, reports its result to the parent, and remains available for review or revisions. For coding tasks that can run in parallel, fork can give each child an isolated Git worktree on its own branch.<br>trace — gives Pi three deterministic ways to navigate code: outline maps the definitions in a file or directory, def retrieves a complete definition by name, and callers finds call sites to inspect. Tree-sitter provides the syntax trees, while SQLite caches the resulting index. callers is deliberately simple: it finds call-shaped syntax without resolving imports or types, so it also works on incomplete and broken source.<br>scry — lets Pi search the web. Its single web_search tool returns links with enough context for the agent to decide what to open, and can restrict searches to recent results.<br>mine — lets Pi read those pages. Its web_fetch tool opens a URL in Chrome, waits for JavaScript to render, dismisses cookie banners, and extracts the main content as clean markdown. Using a browser rather than a simple HTTP request makes client-rendered sites work too.<br>So far, I haven't needed anything else.<br>Working with and building agents has taught me where they struggle and what kinds of support they need. That knowledge shaped these extensions and has been as valuable as the tools themselves. It has also shaped what I deliberately left out.<br>Explore, then validate<br>Every codebase is unfamiliar to an agent. Before it can make useful changes, it needs orientation: a high-level map of what is defined where, so it can decide what to inspect instead of reconstructing the architecture through a long sequence of grep searches.<br>A language server appears to be the obvious solution,...