You understand context management deeply. You know the difference between a workflow that makes LLM calls and a true agent loop with tool calling. You know how to start with a smart model and move to cheaper, faster ones without relying on prompt hacks, “CRITICAL:” advisories, or endless lists of dos and don’ts.
You understand what belongs in tools and APIs versus what belongs in natural language. Designing that boundary should be a fixation for you.
You also understand what is structural and what is in the domain of tone, framing, or model “dark magic.” You care about the headspace the model is operating in, the quality of the user experience, and whether the product actually works for confused real people.
Despite working on agents, you are not in “Gas Town.” You do not believe every problem requires a meta-harness, and you do not outsource your judgment to chatbots. You know when to escalate to MLEs if a problem likely requires fine-tuning or more advanced methods.
You care deeply about user outcomes. You measure how your experiments are doing, proactively solve quality problems, and have the frustration tolerance required for ambiguous chatbot engineering.