Foundations — Detail & Examples

Part of Agentic Architecture · overview: domain-1.html

In this page

A worked example distinguishing conversational, workflow, and agent architectures on one running task, a turn-by-turn trace of the action loop, and a decision guide for which pattern to reach for.

1.1aThe three-way distinction, with one running example

Take a single task — "handle this inbox" — and see how it changes shape depending on the architecture:

Same task, three architectures

Conversational: a person asks "summarize my unread emails" and the assistant replies. No independent action, no loop — it answers and stops.
Workflow: every incoming email is automatically labeled by sender domain, forwarded to the right folder, and archived after 30 days. The steps never change no matter what the email says.
Agent: the system reads each new email, decides whether it needs a reply, drafts one if so, checks the calendar for scheduling conflicts if a meeting is proposed, and only sends after checking its draft against the sender's prior tone. Each step's outcome changes what the next step does.

1.1bThe action loop, traced

The agent version above is really one loop, repeated: act (read the next email) → observe (what does it say, who's it from) → decide (reply, ignore, or escalate) → act again. Here's what that looks like turn by turn:

ACTRead email from j.rivera@acme.com, subject "Re: Thursday sync"

OBSERVESender is proposing 2pm Thursday; calendar shows a conflict at 2pm

DECIDECannot accept as proposed — check calendar for open slots before drafting a reply

ACTQuery calendar for open 30-min slots Thursday afternoon

OBSERVE3:30pm and 4pm are open

DECIDEDraft a reply proposing 3:30pm; hold for send until reviewed (see Domain 1.4)

1.1cWhen each pattern is the right call

Decision guide

If the exact steps are the same regardless of input — use a workflow. It's cheaper, faster, and its behavior is fully predictable before you ever run it.
If the task is answered from context already in hand, with no need to independently act on anything — a conversational turn is enough.
Reach for an agent only when the right next step genuinely depends on what an earlier step returned, and that dependency can't be hard-coded in advance.

The exam consistently rewards picking the cheapest architecture that satisfies the requirement — agentic design is powerful but is not the default-correct answer just because it's more capable.

Agentic Architecture Overview 1.2 Task Decomposition →

Claude Certified Architect — Study NotesDWG 1.1