Agent Basics
What agents add beyond chat models, and the nine product questions that define a working agent system.
“Agent Basics for Product Managers” part one
Series home · Next: When to Use an Agent
If we give a model a long role description and tell it to think like a senior product manager, have we built an agent?
Not yet. We changed how the model responds, but we did not give it a system in which it can keep working.
From one response to one task
A chat model usually receives input and generates output. An agent is expected to pursue a task:
Understand the goal → find missing information → choose an action
→ call a tool → read the result → revise the plan
→ continue or request approval → satisfy the completion criteriaThe model provides understanding and judgment. The runtime, tools, state, and permission system turn those judgments into controlled progress.
Three products built around the same PRD request
“Create a PRD from these meeting notes” can produce three different products.
- A generative feature returns a document.
- A copilot proposes an outline and lets the user edit each section.
- An agent workspace identifies the current requirement stage, checks for missing goals and metrics, asks targeted questions, drafts with evidence labels, runs a review, and proposes a versioned update for approval.
The third is not more agentic simply because it makes more model calls. It manages goals, process, state, evidence, actions, and responsibility.
Nine parts of an agent
| Part | Product question |
|---|---|
| Goal | What counts as complete? |
| Model | Which understanding and judgments belong to the model? |
| Instructions | Which principles must always hold? |
| Context | What information is needed now? |
| Tools | What can the system read, change, or send? |
| Workflow | Which steps and gates must remain deterministic? |
| State / Memory | How does work resume, and what may be reused later? |
| Permissions | Which actions require whose approval? |
| Evaluation | How do we prove the result and process are acceptable? |
No first version needs every sophisticated mechanism, but a “super prompt” cannot answer these questions by itself.
Product exercise
Complete this sentence:
The user delegates the goal of ______. The system may read ______ and act on ______. It must pause and ask ______ for approval when ______. The task is complete only when ______.
If the design still says only “the user asks and AI answers,” it may be a useful generative feature, but it does not need to be called an agent.
An agent is not defined by how well it talks. It is defined by its ability to keep judging and acting toward a goal inside a controlled environment.