Agent Builder's Lesson: One Smart Model Surrounded by Dumb, Reliable Deterministic Parts
Miserable_Donut8718 · reddit · 2026-09-29
After much trial and error, the author's agent setups all converge on one pattern:
- A single genuinely smart reasoning step decides what to do; fetching, parsing, and final write-up are handled by boring deterministic tools, including Gamma to turn structured output into clean reports.
- Every time the smart part also handled a boring part, that's where it broke. The principle: minimize where the model has discretion, maximize where it's constrained.
- The author asks fellow agent builders how much of their stack is actually the model versus deterministic glue.
More from coding & agent
- Multi-model coding: routing Opus, GPT, Kimi, DeepSeek in one Codex session — omarsar0 · 2026-09-29
- Hand-drawn-style animation coded directly in HTML5 Canvas by Claude Opus 5.5 — Ror_Fly · 2026-09-29
- Stripe's Kaliski: we're asking our users' AI agents for integration feedback — hazelcough · 2026-09-29
- Last pass with Opus: model handles the final edit and even designs a custom-font ending animation — techhalla · 2026-09-29
- Creator finishes an ad in 2 hours using Opus 5.5 plus Magnific MCP, sharing the full workflow and prompts — techhalla · 2026-09-29
- LangChain Podcast: Prefect Team on the New MCP Spec, FastMCP, Jev and Evals — LangChain · 2026-09-29