Thinking like an architect: cheatsheet

The four parts

Part	Role	Watch for
Model	Reads, reasons, writes	Capable and fallible at once
Tools	Functions the model can call	Each one widens what can go wrong
Loop	Act, observe, decide again	Needs a reason to stop
Context window	Everything the model knows right now	Finite; attention is a budget

Spine of the track: architecture is deciding where judgment lives.

Workflow vs agent

	Workflow	Agent
Path	Predefined code paths	Chooses its own route
Strength	Repeatable, debuggable, cheap	Flexible on open-ended work
Weakness	Rigid	Harder to predict, test, afford
Use when	Steps hold for almost every case	Steps depend on what the work uncovers

Default: the simplest arrangement that fits. Escalate to an agent only when the path cannot be known in advance.

The three trade-offs

Trade-off	Question	Rule of thumb
Decide vs enforce	Model judgment or code guarantee?	Must-hold rules go in structure, not prompts
One vs many	Single agent or a team?	Start with one; split when the work splits
Carry vs fetch	In context or retrieved on demand?	Curate the desk; do not pile it

The architect’s questions

Ask	Because
Does this need an agent at all?	Most tasks have a predictable path
Which rules must hold every time?	Those cannot live only in a prompt
Where are the seams?	Every agent boundary can lose context
What earns a place in context?	Everything on the desk competes for attention
What happens when a step fails?	Products are judged on their worst day

Pitfalls

Pitfall	Correction
Agent because agents impress	Fit over sophistication
Guarantees in the prompt	Structure enforces; prompts request
More context = more knowledge	Attention is a budget; curate
Team as the advanced option	Every seam must justify itself

One-liners

Line	Meaning
An instruction is a request; code is a guarantee	Sort rules into should-usually vs must-hold
A demo exercises the model; a product exercises the architecture	Demo success predicts little
”This does not need an agent”	The most senior sentence in the field