Everyone's agent setup is a patchwork of models
The Ask HN on current agentic setups reads like a catalog of improvisation. One person dropped Cursor and Codex for the Claude IDE after Opus 5.5 and deleted all their skills. Others run Pi with GPT 5.6 models, DeepSeek Flash through the API, Claude Code inside OpenShell, OpenClaw for do-X-if-Y automations, or stacks like Opencode Go, Mistral Pro and Kimi at around 100 euros a month. Some run two $100 Anthropic accounts plus a Gemini one. A separate thread from a 60-year-old says Claude Code re-ignited his passion for building, and others in their 50s say it gave them the guts to go solo-founder.
Two things repeat. People pay for multiple subscriptions and swap models monthly, and they prefer a thin harness over an elaborate one. The Anthropic move to stop Claude Code subscriptions from working with OpenClaw, with commenters saying it feels like a money grab, shows the vendors are trying to control where the tokens get spent.
The counterpoint from one commenter: after using Claude instead of coding by hand for two days, the work felt unfulfilling, like getting an A on a test you did not take.
So what?
Your users are model-agnostic and will leave whichever tool locks them in or throttles them. If you build agent tooling, support multiple backends and keep the harness simple. If you depend on a flat-rate subscription for your own workflow, expect the rules to change.