The Right 300 Tokens Beat 100k Noisy Ones

Abstract

Your agent has 100k tokens of context. It still forgets what you told it two messages ago. Prompt engineering taught us to craft the perfect instruction. Context engineering asks a different question: what does your model need to see and what should it never see at all? It’s the shift from writing prompts to designing context. In this talk, we dissect four antipatterns killing your agents and the architectural fixes that actually work: - The Stuffed Prompt – You crammed everything upfront and hoped for the best. Static context doesn’t scale. We explore dynamic loading and context refinement: fetching what’s needed when it’s needed. - The Wrong Tool for the Job – You picked one retrieval method and used it everywhere. RAG isn’t always the answer. We break down four wrong tools (similarity for correctness, static docs for process, LLM for determinism, scripts for reasoning) and the four right tools that match them. - The Goldfish Agent – Your agent forgets everything between sessions. Built-in memory gives you no versioning, no backup, and no portability across agents. We explore external memory you control. - The Vibes Eval – You shipped because it “felt right.” We build eval strategies that prove your context choices work or expose the tokens you’re wasting – and the things to watch out for: bleeding, leaking, and negative scenarios. We use a coding agent to explain these patterns so you learn how they work under the hood – but everything also applies to AI agents in general.