One of the less talked about tradeoffs in agent engineering: where you put dynamic context can affect output quality, cache hit rate, cost and latency. let me explain ⬇️ say you’re building cursor and need to tell the model what’s currently on the user’s screen. you have 3 ways of putting this information in the conversation: 1. 🤖 System prompt: put the state directly in the model’s instructions. best for quality but terrible input cache behaviour for long conversations if it changes constantly. which increases cost and ttft latency. 2. 🛠️ Tool: let the model fetch the state when it needs it. cache friendly and you avoid sending potentially irrelevant context upfront, but now you’re relying on the model to know when to fetch it. And costs the extra tool round trips. 3. 👤Injected user message: Pass the state along with the user’s message. cheap & deterministic, but now the model has to distinguish app state from what the user actually said. system prompt approach uses the least amount of tokens overall which is great to lower hallucinations. But the other two approaches give models full traceability with state changes. there’s no universally correct answer. it can be a combination of all three. you have to consider how often the state changes, how much context you’re sending, whether the model needs it each time, and how sensitive the application is to latency and cost.

People be calling themselves agentic ai engineers without understanding tool use patterns. Steal these patterns for your next project (the last one is goated) ⤵️

my biggest unfair advantage in ai is the knowledge from building agentic systems in production. stuff you won’t learn in any textbooks or courses. I’ll be exposing all these hidden rules in my upcoming tweets so stay tuned 🏃

soft launching crackedlist.dev it’s simple, just add your portfolio website and compete with others to be the #1 in your field you also get: - visibility - feedback - networking listing is completely free before the real launch. Get your early adopter batch now ⚡️