Week 0: The 2x2 You Already Need and Don't Have Yet

Sarah's brief, later in this course, is one sentence: build a customer-support triage system. She reads it once and treats it like one build. It isn't. It's four separate systems stacked under one product name: a live chat widget with a customer waiting mid-conversation, a nightly pipeline nobody watches, an overnight research agent, and Sarah herself iterating in Cursor. Each of those four needs its own answer to two questions most teams never ask explicitly: who's waiting for this to finish, and who decides what happens next, the model or the code. Lesson 05 walks her full decomposition end to end; this lesson is where the two questions that drive it come from.

Gartner puts enterprise agentic AI project cancellation (a project abandoned or shelved after a real build attempt) at roughly 40% by 2027. The retros Gartner's own analysts cite tend to blame the model. That diagnosis is worth doubting on its face: the models involved were usually already strong enough for the task. A more direct explanation is that nobody asked the two questions above before building, so the team answered both by accident, and got an architecture built for a different problem than the one they had.

The mistake this course corrects

If you're mid-build right now, you probably haven't asked either question explicitly. Not because you're careless: almost every agent tutorial teaches exactly one pattern (a chat loop, a tool call, a response) and never mentions that a different answer to "who's waiting?" or "who decides?" would call for a different shape entirely. You picked an architecture the way most people do: by copying the tutorial's shape, not by deciding it fit your problem. That's fine when it happens to fit. It's expensive when it doesn't, and you find out in production instead of on a whiteboard, which is exactly what nearly happened to the layer of Sarah's brief she almost shipped as one architecture instead of four.

The two questions, and why crossing them matters

Neither question has a correct answer in the abstract, and neither is a maturity ladder. They're two independent design choices. Cross them and you get a 2x2: four quadrants, each with its own latency budget, security model, and failure story. This course spends eleven lessons inside that grid.

What you'll be able to do that you can't do now

Who this is for

Comfort with the idea of an LLM calling tools is assumed; this course doesn't re-teach tool-calling mechanics from first principles. No prior Rasa or agent-framework experience is required. If you've taken Tools, Memory, and Multi-Agent Systems, lessons 1-5 here cross-link to it rather than repeat it, but nothing here requires having taken it first.