The four that get conflated most
Keeping these apart settles a surprising number of questions on their own.
| Term | What it is | What it is not |
|---|---|---|
| Context window | The model's active attention space for a single call. | Storage. It resets between calls unless your application refills it. |
| Retrieval | External knowledge fetched at query time and placed into the window. | A replacement for the window. The model still only sees what fits. |
| Persistent application state | Values your system owns and changes: order status, balances, records. | Something the model can reach without a tool call. |
| Memory | Continuity your application stores and replays across turns. | A model capability. There is none between calls. |
The most common architecture error in this group is representing live state as retrievable text. Retrieval is for knowledge that was true yesterday and remains true tomorrow. Tool calls are for values a system owns and changes independently.
The seven primitives
Patterns are assemblies of these rather than categories of their own.
| Primitive | Its one job |
|---|---|
| Tools | Let the model take an action or fetch a result from your code. |
| MCP | Expose a set of tools so several clients reach the same endpoints. |
| Subagents | Hand a scoped sub-task to a separate context, in isolation or in parallel. |
| Hooks | Fire deterministic code on an event to enforce a rule the model cannot skip. |
| Skills | Package a repeatable procedure as a versioned, reusable unit. |
| Agent teams | Coordinate peer agents that each own part of a goal. |
| Dynamic workflows | Assemble the steps of a workflow at run time rather than fixing them. |
Three layers people answer across
| Layer | The question it answers | Chosen for |
|---|---|---|
| Entry point | What does a person or system touch? | The user and the work |
| Build-time interface | What do engineers write against? | The team and the integration |
| Delivery route | Whose infrastructure does the request terminate on? | Compliance and cloud commitments |
Evaluation vocabulary
| Term | Meaning |
|---|---|
| Code-based eval | A deterministic check, milliseconds, near-zero cost. For anything with one right answer. |
| Model-based eval | A judge model scoring qualities needing interpretation. Costs about as much as the call it grades. |
| Human-review eval | Human judgement for high-stakes or novel behavior. Does not scale, so it is reserved. |
| Golden dataset | The inputs an eval runs against, including edge cases and counterexamples. |
Guardrail decision points
A control at one point does nothing for the others.
| Point | Runs | Usually |
|---|---|---|
| Input screening | Before the model call | Model-based, since injection phrasing cannot be enumerated |
| Output screening | Before the response reaches the user | Model-based, since policy compliance needs language understanding |
| Tool-call authorization | Before any side-effecting action | Deterministic, because authorization must be auditable |
If a term on the exam sent you looking and it is not here, tell me in the Q&A on any of the practice courses and I will add it. A word you had to translate under time pressure is exactly what a glossary is for.