Beads vs Gastown: Should You Use External Frameworks for AI Agents?
Analysis of Beads (agent task tracker) and Gastown (agent orchestration). Why building your own agent setup beats using external frameworks.
In our chat, @og_mishgun was hyping up beads (task tracker for agents) and gastown (multi-agent orchestration).
Beads: A Minimal Agent Task Tracker
Beads seemed native and minimalist: just tracks tasks in .jsonl, structures them, decomposes, helps finish things instead of abandoning halfway. Everything stays in your repo, context isn't lost.
Gastown: Orchestration Overkill
Gastown was the opposite: tons of files and roles (a mayor and convoy?). Decided I don't need it.
Why External Frameworks Are a Trap
I'm skeptical about agent frameworks: you need to study them, read their prompts, to use them as designed. It's useful to understand what other people invent (see personal AI).
But instead of "installing" a pile of someone else's files, prompts, and MCPs, I prefer cherry-picking what I actually need. Don't put your mental train on someone else's tracks.
First-Party Is the Best Party
The industry moves fast: Anthropic announced they were inspired by beads and upgraded their todo into a full task tracker. Now you need to rip beads out again — Anthropic's experts with infinite test tokens will build their own tool better.
Bottom line: study other people's tools, but build your agent yourself. You don't need someone else's legacy.
Original post: @danokhlopkov
Dan Okhlopkov — AI agent practitioner. Building tools for TON Blockchain analysis and Telegram automation.