1.4 When to Create Your Own Agent
Key Takeaways
- Create an agent when the same specialised task recurs across people or time; use ordinary Copilot Chat for one-off work.
- An agent packages a fixed knowledge scope, standing instructions, and a consistent output format so every user gets the same quality.
- If a task needs a different source every time it is performed, it is a prompt, not an agent.
- Check the Agent Store before building: adopting an existing agent costs nothing to maintain.
- Agents do not grant new access — each user still sees only the content their own permissions allow.
1.4 When to Create Your Own Agent
Quick Answer: Build an agent when a specialised task repeats — same kind of question, same body of knowledge, same expected output, asked by many people or by one person many times. If the task happens once, or draws on different sources each time, a well-written prompt is the right tool and an agent is overhead.
The blueprint bullet is "understand the use case for creating your own agent". Note the word use case. The exam is not asking how to build one here — that is section 5.2 — it is asking whether you can tell the difference between a problem an agent solves and a problem it does not.
The Four Tests
Run a candidate task through these questions. An agent is justified when the answer is yes to at least three.
| Test | Question | Agent justified when |
|---|---|---|
| Repetition | Will this task be performed many times? | Weekly or more, or by several people |
| Stable knowledge | Do the same sources answer it every time? | A fixed library, site, or document set |
| Consistency matters | Does it matter that everyone gets the same format and rigour? | Output feeds a process, report, or customer |
| Expertise gap | Do the people asking lack the context to prompt well? | Non-specialists querying specialist material |
Worked Examples
- HR benefits questions. Hundreds of employees, one stable policy library, answers must be consistent and correct, and employees have no idea how to prompt for benefits nuance. Four out of four — build an agent.
- Onboarding guidance for a specific product line. Recurring, stable documentation, consistency valuable. Build an agent.
- "Summarise this contract I just received." Different contract every time, no stable knowledge base. Use a prompt, and reference the file.
- A one-off competitive analysis for next week's board meeting. Happens once. Use Copilot Chat, not an agent.
Agent versus Saved Prompt
Both standardise work, and candidates conflate them.
| Saved prompt | Agent | |
|---|---|---|
| What it fixes | The wording of the request | The wording and the knowledge scope and the behaviour |
| Sources | Chosen fresh each run | Attached once, used every time |
| Best for | A repeatable request | A repeatable role |
| Effort to maintain | Minimal | Real — knowledge and instructions need upkeep |
The practical rule: if writing down the perfect prompt solves the problem, you do not need an agent. You need an agent when the prompt alone is not enough because the user also has to know which sources to point at and what good output looks like.
Check the Store First
Before building, look at what already exists. Microsoft ships agents such as Researcher and Analyst, and the Agent Store carries both Microsoft and third-party agents. Adopting one costs nothing to maintain; building your own means owning its knowledge sources and instructions indefinitely. Build only when nothing available covers the specific knowledge your organisation holds.
On the Exam: An agent is not a security mechanism. It does not grant access to anything, and it does not hide anything. Every user querying an agent still sees only what their own permissions allow, so two colleagues can ask the same agent the same question and receive different answers.
The Cost Nobody Plans For
Agents decay. The policy library moves, the instructions stop matching how the team works, the suggested prompts describe a process that changed. An abandoned agent giving stale answers is worse than no agent, because users trust it. Before creating one, name the person who will own it — if no one will, that is a strong signal the task belongs in a shared prompt instead.
A Short Decision Path
Walk a candidate task down this path and the answer usually settles itself.
- Will this happen again? No → use Copilot Chat. Yes → continue.
- Do the same sources answer it every time? No → save a prompt instead. Yes → continue.
- Does something in the Agent Store already do it? Yes → adopt it. No → continue.
- Will someone own it a year from now? No → build a shared prompt. Yes → build the agent.
The fourth question is the one organisations skip, and it is the one that determines whether the agent is still useful in twelve months.
Scale Changes the Answer
The same task can justify different tools at different sizes. One salesperson answering their own pricing questions needs a good prompt. Forty salespeople answering customer pricing questions from a shared rate card need an agent, because the cost of inconsistency has grown while the cost of building has not. When a scenario stresses how many people perform the task, or that answers must not vary, it is pointing at an agent.
Which task best justifies building a custom Copilot agent rather than writing a good prompt?
Two colleagues ask the same custom agent an identical question and receive different answers. What is the most likely explanation?