A team proposes a multi-agent system to reformat invoices, a task with four fixed steps. What do you advise?
Why D is the answer
Use a prompt chain, because the steps are known in advance
Known steps belong in predefined code paths. They cost less and are easier to test. Agents earn their cost when the path depends on what they find.
In detail
- When the steps are known in advance, a prompt chain written in code is the simplest correct design.
- Fixed steps are easier to test and cost less, because no model has to decide what happens next.
- Agents earn their extra cost when the path depends on what is found along the way.
- Add a check between steps in code, so a bad result stops the chain.
Use case
An illustrative scenario, not a real incident.
A finance team reformats invoices with four fixed steps: read, normalize, validate and export. A multi-agent proposal would add planning and hand-offs to nothing that varies. A prompt chain in code does it for less, and each step can be tested.
See it in different ways
flowchart TD
A{"Are the steps known?"} -->|Yes| B{"Does the input type change the path?"}
A -->|No| C["Orchestrator-workers or agent"]
B -->|Yes| D["Routing"]
B -->|No| E["Prompt chaining"]mindmap
root((Workflow patterns))
Fixed path
Prompt chain
One of several paths
Routing
Same task many times
Parallel
Path found at run time
Orchestrator workers
Quality loop
Evaluator optimizerflowchart LR S0["Read invoice"] --> S1["Normalize fields"] --> S2["Validate"] --> S3["Export"]
Key point: Known steps belong in a chain, not in agents.
Why the other options are wrong
- A. Keep the proposal, because more agents give more reliable results at any scale
- More agents add calls, cost and failure points. They do not make a fixed task more reliable.
- B. Use an evaluator-optimizer loop so that each step is reviewed by a second model
- An evaluator loop suits work with clear quality criteria that improves over rounds. Reformatting does not need one on every step.
- C. Replace the steps with one autonomous agent that plans its own order of work
- An autonomous agent adds uncertainty to a task whose order never changes.