Module 3 | Sequential, Parallel & Routing | AI 104

Fixed order

8 min

Pattern A: Sequential

You have broken the work into steps and marked the Dependencies. The question now is how to arrange those steps. Should all work run in a fixed order? The answer depends on the dependencies. A Workflow Pattern is one of a few common shapes for arranging steps.

Sequential: A → B → C → D

A: collect data
B: analyze
C: write a draft
D: review

Text equivalent: four steps in a line; each step waits for the previous one to finish and uses its output.

Sequential fits when

Later steps need the outputs of earlier steps
The order itself matters (for example, review must happen before sending)
The flow is simple and needs to be easy to trace and check

Watch out for

An error in any step affects all later steps, so every step needs a check
Independent work forced into a line wastes time
The order should come from a Dependency, not from habit

Example (simulated)

Weekly feedback report: collect feedback → classify → write the draft report → a person reviews. Each step needs the result of the one before, so it is Sequential.

Check yourself

Which situation fits Sequential best?

Three independent analyses that are combined into one report at the endLater steps use the outputs of earlier steps, such as collect → analyze → write → reviewCompletely different handling flows depending on the message category

Key takeaway: Order should come from a Dependency: if a later step needs an earlier output, use Sequential.

Other Workflow Patterns

7 min

Pattern B: Parallel

Some steps do not need each other's results. Lining them up one after another only makes the whole flow slower. They can run at the same time and be combined at the end.

Parallel: run separately, then Combine

A: prepare the shared input (for example, the same brief)
B, C and D at the same time: financial analysis, policy analysis, market analysis
Combine: merge the results into one
Review: a person reviews the combined result

Text equivalent: A prepares the shared input; then three independent analyses B, C and D run at the same time and do not need each other's results; once all three finish, a Combine step merges them into one report, and a person reviews it.

When to use Parallel, and what to watch
Item	Description	Example (simulated)
Condition 1	The tasks are independent and do not need each other's output	Financial, policy and market analysis each use different data
Condition 2	The results can be combined later	Merged into one overall report
A Combine step is needed	One step must integrate the results and handle contradictions	When two analyses conflict, flag it and hand it to a person
Watch out	Decide what happens when one branch fails	If one analysis lacks data, flag the gap instead of silently skipping it

Check yourself

A report needs three independent analyses, followed by one synthesis step. What is the best design?

Sequential: finish the three analyses in turn, then synthesizeRouting: send each analysis type to a different flowParallel + Combine: run the three analyses at the same time, then merge and synthesize

Key takeaway: Independent work whose results can be merged suits Parallel, and it always needs a Combine step.

8 min

Pattern C: Routing, and a preview of Loop

Sometimes different inputs need different handling. Routing classifies the input first and then sends it down a path according to its category.

Routing: Input → Classify → Route A / B / C

Input: an input arrives
Classify: decide which category it belongs to
Route A, Route B or Route C: follow the flow for that category
Fallback: when the category is unclear or confidence is low, hand it to a person

Text equivalent: the input first passes a classification step, then splits into three paths of which only one is taken; if it cannot be classified, it takes the Fallback path to a person.

Common bases for Routing (simulated examples)
Basis	Example paths
Customer issue category	Billing → billing flow; technical → technical flow; general → general reply
Document type	Contract → legal review; invoice → finance check; application form → administrative handling
Risk level	Low risk → handled automatically; high risk → a person approves
Approval path	Small amount → unit head; large amount → escalate one level up

Optional: Loop (a bridge only; deeper coverage belongs to later courses)

Execute: run once
Check: check the result
Good enough? Yes → Finish
No → Revise, then return to Check

Info

A Loop is a repeated execute, check, revise-if-not-good-enough structure. It needs clear checking criteria and a stop condition, otherwise it can repeat forever. Module 4 introduces the basic Agent Loop; deeper Planning belongs to later courses.

Check yourself

Purchase requests are split by amount: below a threshold they are handled automatically, above it they go to a manager for approval. Which Pattern is this?

Routing: inputs are sent down different paths by a conditionSequential: every request goes through the same steps in orderLoop: the same request is processed again and again

Key takeaway: Routing chooses a path by condition and keeps a Fallback. A Loop repeats execute, check, revise and needs a stop condition.

Practice and apply

7 min

Practice: Choose the Workflow Pattern

Below are 8 simulated scenarios. Choose Sequential, Parallel, Routing or Loop for each and explain why in one sentence. Some cases can reasonably combine two Patterns; what matters is that you can explain your reasoning. Please do not enter real company secrets, client personal data or any unauthorized sensitive information in this exercise.

8 scenarios (decide for yourself)
Case	Scenario	Think about
1	Collect data → analyze → write the report → review	Do later steps need earlier outputs?
2	Financial, policy and market analyses are independent and merged into one report at the end	Do they need each other's results?
3	Support messages are sent to different handling flows by category (billing, technical, general)	By what condition is the path chosen?
4	A draft is checked against a checklist; if it fails, revise and check again, at most 3 times, then hand it to a person if it still fails	When should it stop?
5	A document is first identified as a contract, invoice or application form, then follows its own review flow	Classify first, or execute first?
6	Query three different data sources at the same time, then organize the results into one summary table	Do the queries depend on each other?
7	Student report: check format → grade → write comments → teacher confirms	Does the order matter?
8	Requests by risk level: low risk is handled automatically, high risk goes to human approval; once approved, create the record and then notify the applicant, in order	Which Patterns does this scenario use?

Sequential: later steps depend on earlier outputs; order matters
Parallel: independent work whose results can be merged later
Routing: choose a path by condition (category, risk, type)
Loop: execute, check, revise until the standard is met or a stop condition is reached

Go to the Practice page: Choose the Workflow Pattern

Check yourself

Case 4 uses a Loop to revise a draft repeatedly. Which design is the safest?

Keep repeating until the AI itself feels satisfiedSet clear checking criteria and a maximum number of repeats, and hand it to a person if it still fails after thatRepeat only once and send the result regardless of quality

Key takeaway: Look at Dependencies to choose Sequential or Parallel, at conditions for Routing, and at "needs repeated revision" for Loop; real flows often combine several.

5 min

Apply: put the Workflow Pattern on the Canvas

Done

Look at the steps you broke out in Module 2 and ask: which must run in order, which can run at the same time, where should the flow branch by condition, and where does it need repeated revision?

What to add to the Canvas: Execution Order
Question	Write your answer as	Example (simulated)
What is the main order?	The Sequential steps and their order	S1 → S2 → S3 → S4 → S5
Which steps can run at the same time?	The Parallel steps, and where the Combine happens	S2 classify and S3 find follow-ups can run separately, combined in S4
Where does it branch by condition?	The Routing condition, each path, and the Fallback	Different follow-up flows by feedback category; unclear category goes to a person
Where does it need repeated revision?	The Loop checking criteria and maximum repeats (if any)	Revise the draft against a checklist, at most 2 times, then hand to a person if it still fails

Module 2: Task decomposition

Break out steps, Inputs and Outputs
Mark the Dependencies

Module 3: Workflow Patterns

Choose Sequential or Parallel from the Dependencies
Add Routing and Loop where needed

Human-AI-Human: a person chooses the shape of the flow and the branch conditions, the AI executes along the flow, and a person reviews and stays responsible for the result. The next module asks: which steps can be fixed in advance, and which need the agent to judge?

Continue to Module 4: Decision Points & Agent Loop

Key takeaway: Write the chosen Workflow Pattern and your reason for it into the Execution Order of the Canvas.

Sequential, Parallel & Routing

Learning objectives

Fixed order

Pattern A: Sequential

Sequential fits when

Watch out for

Other Workflow Patterns

Pattern B: Parallel

Pattern C: Routing, and a preview of Loop

Practice and apply

Practice: Choose the Workflow Pattern

Apply: put the Workflow Pattern on the Canvas

Module 2: Task decomposition

Module 3: Workflow Patterns

What next?