Rewriting my agents: most steps are plain code, LLM only needed for one judgment call
Charming-Hawk-5760 · reddit · 2026-10-05
The author runs a few personal agents (server monitoring, site caretaking, Telegram replies) and realized most of the work doesn't need an LLM. After rewriting: plain code drives the job, and the model is called only for the one step of reading text and making a decision, with answers validated against a schema. It's also much cheaper — steps that need no reasoning are free. The author asks whether others found the same, or if there are cases where letting the model pick every step actually paid off.
More from coding & agent
- Amazon (2B) and Cloudflare (9B/27B) ship open-source decision models rivaling Jev — mark_k · 2026-10-05
- Autonomous agents only pay off where value justifies engineer escalation, dev argues — Pavel_Asparagus · 2026-10-05
- Users report agents like Codex feel noticeably slower than three months ago — koltregaskes · 2026-10-05
- Building After Effects motion graphics engines with Opus 5.5 — notiansans · 2026-10-05
- Local inference in practice: mining repos, AI chat logs and media libraries with a tiny model — natesiggard · 2026-10-05
- repOx packs 3k-file repos into LLM prompts in ~14ms, stripping 30k-token lockfile waste — Wvdy_CC · 2026-10-05