A non-programmer built a multi-agent prototype for petrochemical trading and learned why agents must be modular
Sphere_Project · reddit · 2026-07-25
A non-programmer recounts a fourth attempt at getting into IT, and how a real job problem pushed him into building an AI system.
He wanted a bot that would watch Platts petrochemical prices, combine them with seasonality, holidays, weather, jet-fuel demand and news, and then alert him when to buy Brent-related benchmarks. Building it exposed five practical problems:
- Data access: Platts quotes are not free, so scraping and reliable polling became necessary.
- Memory: the connected AI kept forgetting context, making multi-year pattern comparison hard.
- Multi-source fusion: weather, jet-fuel data and news all came in different formats and had to be connected meaningfully.
- Failure isolation: a single input→analysis→output chain was too brittle, so the system had to be split into independent pieces.
- Incomplete first version: the original bot never worked as intended, but the architecture lessons became the seed of his current multi-agent setup.
The post is less about a finished product and more about the engineering reality of turning an agent idea into a resilient system.
More from coding & agent
- Bio is building a domain-specific research agent and surveying researchers first — melnykowycz · 2026-07-25
- Kimi K3 lands in NEO and beats GLM 5.2 on a real AI engineering task — Nilofer_tweets · 2026-07-25
- Open-source TokenShield cuts off CrewAI retry loops before token burn explodes — bulleykebaal · 2026-07-25
- OpenAI reportedly ships Presence, an enterprise agent product that resolves 75% of support issues — dl_weekly · 2026-07-25
- Browser agent benchmark shows diff-based page memory cuts token growth 37% — Desperate_Title1595 · 2026-07-25
- Hermes pitch claims persistent memory and skill generation can make agents self-improving — Teknium · 2026-07-25