When to Use an Agent
Use task shape, context certainty, and workflow certainty to choose the simplest effective AI architecture.
“Agent Basics for Product Managers” part two
Previous: Agent Basics · Series home · Next: How Agent Runs Work
The easiest mistake in agent product design is not choosing the wrong model. It is giving a system too much freedom for a task that could have been completed deterministically.
Ask four questions first
- Do the required steps change with the input or intermediate results?
- Must the system access external information or take real actions?
- Must progress survive multiple turns, pages, or periods of time?
- If the system is wrong, can the user detect, stop, and recover from it?
The first two determine whether action-oriented intelligence is useful. The last two determine how much state and control the product needs.
Start with the simplest option
| Task shape | Prefer | Example |
|---|---|---|
| One act of understanding or generation | Single model call | Summarization, classification, extraction |
| Multiple fixed steps | Workflow | Parsing documents, cleaning transactions |
| The path changes and tools must be selected | Agent + tools | Research, requirement discovery, code repair |
| A fixed main path contains open decisions | Workflow + agent | Contract review, BRD to PRD |
Multi-agent is not a fifth default. It is an organizational pattern that may become useful later.
Two variables determine where freedom belongs
- Context certainty: Is the required information complete, structured, and trustworthy?
- Workflow certainty: Are the steps predictable enough to define in advance?
When context is uncertain, improve sources, retrieval, clarification, and evidence. When workflow is uncertain, design planning, tool selection, stopping conditions, and review. When both are certain, reduce model autonomy.
For a PRD task, consistent notes with complete fields and a fixed template may need only a workflow with model-assisted extraction. An agent becomes valuable when the system must detect missing goals, ask different questions, consult related material, and decide when the requirement can advance.
Product exercise
Score context certainty and workflow certainty from one to five for a candidate feature. Name the main uncertainty, then decide whether data quality, a workflow, agent judgment, or a human should resolve it.
Begin with the minimum autonomy that solves the problem. Expand it only in response to observed failures and evidence.