Software Research

Most of my projects have a deliverable. This one doesn’t — on purpose.

Software Research is the space where I’m allowed to spend a weekend getting a tool running, poking at it, deciding I don’t like it, and walking away. Without that being a failure. The output is not a product. The output is a short document per topic: what it is, what it’s good for, what it’s bad at, and whether I’m going to adopt it, drop it, or shelve it for later.

The rule is bias-to-action: no living in the documentation. Every ticket closes with something that runs — a repo, a command, a demo. Otherwise it wasn’t really research.

What’s already come through here

• Linear Agents — native agents inside a Linear workspace. Research ticket became a hands-on, hands-on became hello-agent (end-to-end, deployed at hello-agent.gabi.tv), which became coder-agent — a real coding agent with Claude Agent SDK + GitHub App + per-session workspaces that clones repos, edits code, and opens PRs. Live at coder-agent.gabi.tv. Both agents were graduated out of this project into their own.
• Aider + OpenRouter — prompted by GitHub Copilot being cancelled on 2026-04-18 (effective 2026-05-10). Terminal-first coding agent, multi-model via OpenRouter, no lock-in. Decision pending against Claude Code / Continue / Cursor.

What’s in the queue

A radar of candidates that don’t yet deserve their own ticket: Continue (VSCode autocomplete via OpenRouter), Zed editor, the official MCP SDK, self-hosted OpenRouter alternatives. They get promoted out of the backlog when one of them earns attention — or stay here forever if they don’t.

The counterintuitive trick: having a sanctioned place to not commit to things is what makes the other projects possible.