Part 4 of 9 · 1 min read
The ladder
Five common shapes of AI software, from least to most freedom for the model.
Think of a ladder. The higher you climb, the more the model decides for itself. Each step up gives more power and costs more control.
| Shape | Who chooses the steps? | Plain example |
|---|---|---|
| Chatbot | Nobody. One question, one answer. | "Explain what a refund is." |
| Model with tools | The model picks a tool for one request. | "What is the status of order 4471?" |
| Workflow | Your code fixes the order of steps. The model does the thinking inside each step. | Always: look up the order, check the policy, write the reply. |
| Agent | The model chooses the next step each time, and when to stop. | "Sort out this refund." The model finds its own path. |
| Multi-agent system | Several agents, each with its own job, pass work to each other. | One agent gathers facts, another writes, a third checks. |
Workflow versus agent
In a workflow, your code fixes the sequence of steps, and the model does the thinking inside each step. The same input follows the same path. In an agent, the model chooses the next action and decides when to stop.
Example: "look up the order, then check the policy, then reply" is a workflow. "Work out what this customer needs and do it" is an agent.
A workflow is usually cheaper, faster and easier to test. Choose an agent only when you truly cannot list the steps in advance. Level 3, module 4 shows how to decide. Many good systems are a workflow with one agent placed only at the step that needs freedom.
Check yourself:
- Which rung of the ladder is "always run these three steps in this order"?