HN's Agentic Setups: Cheap Open Models Do the Work, Frontier Models as Second Opinion
michaelTM_ai · reddit · 2026-11-05
After reading 45 of 475 replies to HN's "what's your agentic setup" thread, recurring patterns emerge: people run cheap open models (DeepSeek mentioned more than any frontier model, plus GLM and Kimi) at home via a $10 opencode plan or openrouter pay-as-you-go; frontier models like Opus are kept as a second opinion for review, architecture, and the tricky bits — Revolut's 2k-engineer approach at personal scale. Harnesses are shrinking (pi, opencode; top comment deleted all skills and just describes "what done looks like"). Almost nobody trusts all-in-one agent platforms like openclaw or hermes — orchestration is still a human with 3-5 terminal panes, with one person typing junk every 30-45 min to keep caches warm. Best reply: "people bragging about multiple subscriptions without saying what they use them for."
More from coding & agent
- Team's 24/7 infra-monitoring agent spots anomalies 'before the fire,' user says — antoine_chaffin · 2026-11-06
- Delvetown: a human-agent "post-scarcity" web community, plus agents emailing researchers unbidden — Miles_Brundage · 2026-11-06
- Every team demos Codex and Claude workflows for running AI as your OS — every · 2026-11-06
- Hyperagent launches Sizzle, an agent that reads PRs and auto-produces polished demo videos — aakashgupta · 2026-11-06
- Vercel publishes CLI design guidelines for humans and AI agents — JohnPhamous · 2026-11-06
- Dev builds a sculpting tool for Apple Vision Pro using Codex — OpenAIDevs · 2026-11-06