Agent engineering lessons: the model is the smallest, most swappable part
alex_verem · x · 2026-09-14
Responding to Garry Tan's "become a system of record or a domain-specific harness" take, the author distills key principles from engineering agent harnesses:
- Garbage in, garbage out: better models won't fix bad context; information quality determines output quality
- Hardcode fixed rules: keep calculations and critical metrics in deterministic tools so the model never hallucinates exact numbers
- Keep humans accountable: agents advise, but a designated person approves actions before execution
- Audit every step: detailed logs let you trace failures to the exact prompt or tool and patch permanently
- Build for interchangeable models: with solid architecture, swapping the underlying model is just a config change
The model gets the credit, but the infrastructure delivers the results.
More from coding & agent
- Dev hails opencode2 + DeepSeek V4.1 Flash as an insanely fast coding combo — Scobleizer · 2026-09-14
- Builder of 400+ production agents shares a decision tree for agent frameworks — MaryamMiradi · 2026-09-14
- Gemini 3.8 Flash with Antigravity draws elaborate diagrams right in the terminal — doodlestein · 2026-09-14
- $100/month coding agent beats buying a $10 app, quips AI researcher Yoav Goldberg — yoavgo · 2026-09-14
- SkySynth synthesizes formally verified systems: KV stores 2.3x faster than Redis — CShorten30 · 2026-09-14
- browser-use Open-Sources Agency: An Agent That Finds Work for You, You Just Approve — TheMoonMidas · 2026-09-14