Module 3 · Lesson 5 of 8

We now have several useful steps: find venues, check availability, calculate costs, compare options, and request approval. The next problem is their relationship.

Calculating a final total depends on receiving prices. Booking depends on approval. Venue research and catering research may be able to happen together.

A graph is a map of connections

A graph contains nodes and edges. In a workflow graph, nodes represent work and edges represent possible transitions or dependencies. An arrow indicates a direction.

The drawing makes the order visible, but it does not by itself run anything. Software must implement what the nodes do and when the connections can be followed.

Branches express choices

After checking a venue, the work can branch. An unavailable venue goes back to the search. An available venue moves to cost checking. A price above the limit goes to another option or back to the organizer.

A branch can use an explicit condition such as “capacity is at least 40.” It can also use a model's classification or decision. Conditional routing is not automatically AI behavior; ordinary programs branch too.

Parallel work needs a meeting point

Suppose venue research and catering research run at the same time. A join is where the needed results come together. Before comparing complete packages, the application must know whether it has received both results or whether one branch failed.

Starting tasks together is only half the design. Someone must define when their combined output is ready to use.

State is the current record of the work

State is the information the workflow carries as it runs. Think of a shared event worksheet:

Guests: 40
Budget: $2,000
Venue quote: $800
Catering quote: $1,000
Other costs: not checked yet
Approval: pending
Booking confirmation: none

The “not checked yet” entry matters. An absent cost is not the same as zero. The model should not describe a complete budget before those remaining costs are understood.

Different nodes can read and update different fields. If two workers update the same information, the application needs a rule for combining changes. Otherwise, a newer result might be silently overwritten.

For a framework example of state, nodes, edges, and update rules, see the LangGraph Graph API.

There is more than one kind of graph

A knowledge graph could record “Riverside Hall is located in the town center” and “Riverside Hall offers catering.” Its connections describe relationships between facts and entities.

Our workflow graph connects activities. A system can use both, but drawing the workflow does not require a graph database. Ordinary application code can represent it.

State tells us what we currently know. Nodes do work. Edges describe how that work connects. This gives us the vocabulary to design a larger system without treating every step as an agent.