Reddit dev gates agent links with a hand-tuned keyword score instead of the LLM
Most-Agent-7566 · reddit · 2026-09-23
A developer describes a two-layer architecture for his Reddit-posting agent: an LLM reads threads and writes comments, but whether a comment is allowed to carry a link is decided by a deterministic gate with no LLM involvement.
The gate: count "strong" intent terms (exact product names) and "weak" terms (paraphrases, worth half) in the original post; if the weighted score clears a threshold and no anti-intent terms appear, the link survives. About ten hard boolean checks (thread allows links? mod already declined?) supplement it — none consult a model.
Why not trust the LLM: the author argues a model will rationalize "yes, this person wants what I sell" because agreeing with itself is free, so the gate is deliberately dumb.
Open questions: the gate has caught false positives but almost certainly misses intent phrased outside the lists, and it's been finding nothing for weeks — correctly or not, he can't tell. His core worry: when does a deterministic gate stop being a safety rail and become a second, worse model that nobody retrains, and is there a shape that keeps the "don't let the model grade itself" instinct without freezing into stale lists?
More from coding & agent
- Studio Jadu tests its AI assistant with annoying, impatient synthetic users — andimarafioti · 2026-09-23
- Fine-tuning a 194M GLiNER2 tag classifier on HF Jobs for ~$1.50 lifts accuracy from 10% to 69% — vanstriendaniel · 2026-09-23
- Dev builds 3D Spider-Man game from scratch with Opus 5.5 in Three.js — TawohAwa · 2026-09-23
- Alibaba Launches Qwen Intelligence With 3 SOTA Mobile Agents, 90% End-to-End Success Rate — Alibaba_Qwen · 2026-09-23
- Deleting Three Agents for 40 Lines of Python Cut Latency 9s to 800ms and Cost 75% — Prestigious_Style267 · 2026-09-23
- Automation vs Agentic AI: Most 'Agents' Are Just Scripts, LLM Only Fills the Gaps — forevergeeks · 2026-09-23