A custom agent gives Copilot a defined role, instructions and knowledge. More advanced agents can also use connectors, workflows and business actions.
Simple agents may be included; tenant data, advanced actions and Copilot Studio can require paid or metered capacity.
Four decisions define an agent
👥AudienceWho will use the agent and for which job?
📚KnowledgeWhich approved sources should shape its answers?
🧭InstructionsHow should it reason, respond and handle uncertainty?
⚙️ActionsWhat may it do beyond giving an answer?
Use an agent for a repeatable specialist role
📘
Knowledge guidance
Create a policy, product or process agent that answers from selected approved sources and explains when those sources do not contain an answer.
🧑🏫
Role-specific support
Build a manager coach, onboarding assistant or service guide that follows a consistent framework for a defined audience.
⚙️
Process assistance
Guide users through standard steps or, in a more advanced agent, invoke a workflow or connected business action.
Agent Builder or Copilot Studio?
Choose
When it fits
What it provides
Agent Builder
You need a straightforward agent with instructions and selected knowledge.
A quick natural-language or template-based creation path inside Copilot.
Copilot Studio
You need richer data, connectors, actions, workflows, governance or deployment.
A full maker platform for more advanced business agents.
Cowork Skill
You want Cowork to reuse one execution method.
Instructions for how Cowork handles a type of task.
Copilot Notebook
You want a curated source workspace, not a separate assistant.
Persistent reasoning over selected references.
Start with the simplest agent that works
Choose one audienceName the people and the specific problem the agent will handle.
Describe the jobUnder Agents, choose New Agent and explain the desired behavior.
Add approved knowledgeUse a small source set that is current and authoritative.
Test before sharingTry missing information, incorrect assumptions and access failures.
Example agent definition
Create an agent for new sales managers in Europe. It should answer questions using the approved manager handbook, compensation policy and regional HR guidance. It should explain which source supports each answer. If the sources conflict or do not cover the question, it must say so and direct the user to HR rather than inventing a policy.
💡
Key insight
The most important agent design choices are its knowledge boundary and action permissions—not its name or personality.
Knowledge and actions create different risks
📚KnowledgeDetermines what the agent can use to answer.
💬InstructionsDetermine how the agent applies that knowledge.
⚙️ActionsDetermine what the agent can change in other systems.
🛡️GovernanceDetermines who can build, use, monitor and update it.
Do not add actions when answers are enough
📚
Knowledge-grounded agent
It answers or guides from defined sources. This is sufficient for many policy, product and onboarding use cases and keeps the risk surface smaller.
⚙️
Action-capable agent
It can invoke a workflow, connector or service. That can save steps, but it needs explicit permissions, validation, approvals and failure handling.
Design what happens when knowledge fails
🕳️
The source is silent
Specify whether the agent should ask for more context, give a general answer with a warning or direct the user to a human owner.
⚖️
Sources conflict
Tell the agent which source has priority or require it to show the conflict instead of merging incompatible guidance.
🕰️
The source may be stale
Assign an owner and review cycle. An agent can answer consistently from outdated knowledge, which makes maintenance essential.
Licensing follows what the agent uses
Agent design
Typical access pattern
Cost consideration
Instructions / public sites
A lightweight declarative agent without shared tenant data.
Can be available through included Copilot Chat scenarios.
Shared tenant knowledge
SharePoint or connected organizational data for an unlicensed user.
Usage-based Copilot Studio billing may apply.
Advanced actions
Connectors, workflows, richer orchestration and enterprise deployment.
Copilot Studio capacity, metering and governance become relevant.
Test behavior, not only good answers
Correct caseConfirm the answer is useful when the approved source is clear.
Missing caseCheck that the agent does not invent policy or certainty.
Permission caseConfirm it fails safely when a user cannot access a source or action.
Action caseVerify inputs, approvals, duplicate handling and recovery after failure.
Move to Copilot Studio for a reason
Use the more advanced platform when the agent genuinely needs richer connectors, actions, analytics, lifecycle control or deployment—not simply because the agent is important.
🛡️
Important limitation
An agent is only as reliable as its knowledge, permissions and maintenance. Sharing it more widely increases the need for ownership, testing and change control.