@betashop.eth

Been meaning to share this with people building in the agent space. We've been running autonomous AI agents in production managing real user funds — $100M+ notional volume, thousands of trades, iterating daily. The speed of learning when real money is on the line is unlike anything else. The biggest thing we learned: we were routing 75% of our agent workload through the LLM that didn't need to be there. RSI checks, trailing stop math, risk threshold monitoring, state management — all deterministic, all better handled by code. When we pulled it out, agents got 4x cheaper, more reliable, and dramatically faster to build. The framework we landed on: scripts generate facts (cheap, fast, deterministic). Plugins handle shared operations (reusable across agents). The LLM handles judgment — and only judgment. If you're building AI skills or agent workflows, this might save you some of the expensive lessons we learned at speed: https://x.com/betashop/status/2029595926227210292?s=20 The TLDR principle: LLMs should think, not calculate. Every token spent on arithmetic is a token wasted. The best agents aren't the ones where AI does everything — they're the ones where AI does only what AI is uniquely good at. Curious how others are thinking about this split. The "route everything through the LLM" default is tempting because it works — it's just expensive, slow, and fragile at scale.

3 replies

7 recasts

44 reactions