Workflows Before Agents
Choose between a fixed workflow, a single agent, and a multi-agent system based on how predictable the work is and what failure costs.
Objective
By the end of this lesson, you can place a task on the spectrum from fixed workflow to multi-agent system, and justify the simplest design that does the job.
A spectrum, not a switch
"Agent" is used loosely. As an architect, be precise. There are three common designs:
- Fixed workflow. Code decides the steps. The model fills in parts, such as classifying a message or drafting a reply. Example: classify request, look up matching program, draft reply, send to staff for review.
- Single agent. A model runs in a loop. It picks a tool, reads the result, and decides the next step until it finishes or hits a limit.
- Multi-agent system. Several model loops with different roles hand work to each other. Example: one plans, one researches, one checks.
Each step to the right adds flexibility. It also adds unpredictability, cost, delay, and new ways to fail. Each step is harder to test, because the path changes from run to run.
The rule of thumb
If you can draw the steps, code the steps. Use a fixed workflow whenever the path is known. Most intake, routing, and summarizing tasks fit here.
Reach for a single agent only when the right next step truly depends on what was found, and you cannot list the paths ahead of time. Research across many sources is a fair example.
Reach for multiple agents only when the work splits into parts that need different tools, permissions, or context, and one loop gets confused holding it all. Even then, a single agent with good tools is often enough.
| Question | Points to workflow | Points to agent |
|---|---|---|
| Can you list the steps ahead of time? | Yes | No |
| Must every run be auditable step by step? | Yes | Harder |
| Is a wrong step costly or hard to undo? | Yes | Only with strong gates |
| Does the task need to adapt to what it finds? | Rarely | Often |
Common workflow shapes
These patterns cover most real systems without a free-running agent:
- Chain: step A's output feeds step B. Add checks between steps.
- Router: classify the input, then send it down one of several fixed paths.
- Parallel: run several model calls at once, then combine. Useful for checking one answer several ways.
- Evaluator loop: one call drafts, another checks against rules, and the draft is revised a fixed number of times at most.
Quiz: Pick the design
Think it through
Place your use case
Go deeper: Automating Administrative Workflows vs. Creative Workflows
Discussion
Ask. Answer. Build together.
Loading…