Module 4 · Lesson 8 of 8

Return to the original request: organize a community event for 40 people with a budget of $2,000, and ask before booking.

We can now describe how the system handles that request without treating every part as a mysterious form of AI.

Start by recording the requirements

The application records the date, guest count, budget, and approval requirement. If the date is ambiguous, it asks the organizer. These values become part of the workflow's state.

A model can help interpret the request. Code can check that a guest count is a positive number and that a budget has a currency. Neither operation needs to invent missing preferences.

Research within clear boundaries

A venue agent receives the requirements and access to search and calendar tools. Its harness calls the model, executes permitted requests, and returns results. Its loop continues while useful questions remain and its limits permit more work.

A catering agent may work alongside it if the research is sufficiently independent. Both return evidence and unresolved constraints. The graph specifies when their results are ready to combine.

Use code for the arithmetic

The venue costs $800. Catering is $25 per guest, or $1,000 for 40 people. Delivery and equipment add $150. A function calculates the total:

$800 + ($25 × 40) + $150 = $1,950
Budget: $2,000
Remaining: $50

These are fictional figures for the example. The arithmetic is correct, but a real proposal also needs to establish that the quoted amounts cover the required items. Correct addition cannot repair an incomplete quote.

Present the tradeoff and wait

The coordinator explains that the proposal fits the budget with $50 remaining, lists what is included, and flags any uncertainty. The organizer approves that proposal or asks for changes.

While approval is pending, the workflow pauses. It does not need to keep calling the model to ask whether approval has arrived. The application can resume the work when the organizer responds.

Execute and verify

Before booking, the application checks that approval applies to the current proposal. If price or availability has changed, the graph sends the work back for review.

The booking tool performs the reservation. The system records the confirmation or handles an uncertain result. Only then should it report the booking as confirmed.

Where each idea fits

  • Model: interprets information and produces outputs used by the application.
  • Tool: exposes an operation such as checking availability.
  • Function node: performs a specified procedure such as totaling costs.
  • Agent: uses the model to choose actions toward an assigned goal.
  • Loop: repeats a process using new results.
  • Harness: runs model and tool interactions and manages execution.
  • State: records the current task information.
  • Graph: connects the work through dependencies and transitions.
  • Multi-agent system: coordinates several agents with defined responsibilities.
  • Graph engineering: makes those connections and their behavior deliberate and dependable.

A graph can contain an agent, and an agent can contain a graph

At one level, “research venues” is a single node. Inside it, there may be several steps and a tool loop. Conversely, a larger agent application can use a fixed graph to handle a well-defined part of its task.

A cycle in a graph can represent repetition. But neither a cycle nor several model calls automatically implies multiple agents. These ideas describe different aspects of the system, and their boundaries can overlap.

Try the same questions on another task

Imagine a coding assistant fixing a defect. Reading a file is a tool operation. Running a prescribed test is a defined step. Choosing which part of the code to investigate may be agent work. Repeating an edit after a failed test is a loop. Requiring passing checks before delivery is a workflow condition.

For any AI application, ask: what work is required, which parts need judgment, what evidence allows the next step, and how does the process stop? Those questions let you reason about the architecture before choosing a framework.

Continue with agentic AI architecture, or explore why intelligence and execution authority need separate responsibilities.