Part 5 of 9 · 2 min read
One task, three ways
The same refund request handled by a chatbot, a workflow and an agent, and why more steps means more risk.
The task: a customer writes "Please refund order 4471." The numbers below are made up to show the shape of the idea. They are not measurements.
As a chatbot. One model call. The model can only say "Refunds are available within 30 days." It cannot see order 4471. It may sound helpful, but it has not done anything.
As a workflow. Your code runs three fixed steps. Step 1 looks up the order (a tool). Step 2 checks the refund policy (plain code). Step 3 asks the model to write the reply. That is 3 steps every time, for every customer. It is predictable.
As an agent. The model chooses. For one customer it takes 3 steps. For another, the order was split into two parcels, so it takes 5 steps. The path changes with the situation. That is the strength of an agent, and it is also its risk.
The risk comes from multiplying. Suppose each step works 95 times out of 100. That is the made-up number. A run only succeeds if every step works:
| Steps in the run | Chance the whole run works |
|---|---|
| 3 | 0.95 x 0.95 x 0.95 = about 86% |
| 5 | about 77% |
| 8 | about 66% |
Each step looks reliable on its own. But a longer path means more chances for one step to go wrong. That is why real agents need limits (a maximum number of steps), checks (verify the result) and sometimes a person to approve risky actions. Levels 5, 8, 9 and 10 build exactly these.
Read it as
A run succeeds only if every step succeeds, so the chances multiply. Adding steps does not add risk one step at a time. It compounds. This one idea explains why agent builders keep trajectories short and add retries and checks.
Check yourself:
- If each step works 90 times out of 100, what is the chance that a 3-step run works? (Multiply 0.90 three times.)