February 1, 2026
Simple bot, complex feelings
What I learned building an opinionated and minimal coding agent
Dev builds a “no‑frills” code bot, commenters split on simplicity vs speed
TLDR: A developer built a stripped‑down code helper for control and simplicity, skipping fancy features. Commenters split: DIY minimalism fans cheer, while others argue sub‑bots make it faster and security can be handled with sandboxing. It matters because coders are rebelling against bloated tools and rolling their own.
An engineer got fed up with bloated code helpers and built his own minimal “coding agent” — think: a focused little bot that writes code without the extra buttons, pop‑ups, and flickering spaceship vibes. He wants total control over what the bot sees and does, a clean history you can inspect, and the freedom to swap AI models like toppings on a pizza. The community? Oh, they showed up with energy.
DIY fans cheered, calling it a return to simple, predictable tools. [verdverm] said making your own agent is shockingly doable and the real win comes from crafting great instructions. Meanwhile, power users pushed back: [evalstate] argued that subagents — tiny helper bots that handle side tasks — boost speed, directly poking the author’s “no sub‑agents” stance. Then came the business nerds: [sghiassy] wondered if companies like Anthropic have any moat (aka long‑term advantage), stirring up a “can anyone just build this?” debate. Security folks jumped in too, with [charcircuit] snapping back that you don’t have to kill the internet to protect data — “You can sandbox off the data.”
Bonus laughs: people loved the author’s promise to name it so badly no one can Google it, and yes, the “Claude flickers” meme returned, glowing like a haunted dashboard. Minimalism vs megafeatures — the comments turned it into a cage match.
Key Points
- •The author used LLMs for coding over three years, progressing from ChatGPT and Copilot to Cursor and modern coding agents like Claude Code.
- •Frequent changes and feature bloat in Claude Code disrupted the author’s workflows, motivating a custom solution.
- •The author emphasizes strict context engineering and full transparency of model interactions, criticizing hidden injections in existing harnesses.
- •Self-hosting attempts (local and on DataCrunch) were hindered by tooling incompatibilities, notably with the Vercel AI SDK and tool calling.
- •He built a modular, minimal agent stack (“pi”) including pi‑ai (multi‑provider LLM API) and pi‑coding‑agent, with deliberate exclusions (no plan mode, MCP, background bash, or sub‑agents).