The next bottleneck in AI isn't always the model. Sometimes it's latency. Imagine your agent needs to.. read a document, call a database, search the web, run a tool, then ask the model what to do next. If every step waits for the previous one, a simple task can take seconds or even minutes. This is where things like parallel tool calls, streaming, caching and async execution become important. You don't necessarily need a smarter model. You might just need a better system around it. That's something I like about AI engineering right now. A lot of the hard problems are starting to look very familiar, distributed systems, caching, queues, concurrency, observability. The AI part is new. The engineering problems underneath it aren't.