Copilot Notebooks vs. Agents: Sources, Instructions, and Actions

Copilot notebooks and Copilot agents both use Microsoft's AI, but they do fundamentally different jobs. A notebook is a research space: you upload or link sources, ask questions, and get answers grounded in that specific material. An agent is a task runner: you give it instructions and tools, and it acts across your connected Microsoft 365 services. Choosing between them means deciding whether your session needs a reading room or a worker. Tools like fenbs hold shared decisions and work progress across AI tools, so the context you build in one session is available to whichever tool comes next.

The distinction matters in practice. A notebook cannot send a calendar invite or create a task in Planner. An agent does not naturally confine itself to a set of documents you curated for one question. Using the wrong tool for a job produces either inaction (expecting an agent but using a notebook) or unbounded scope (expecting a notebook but using an agent with broad permissions).

What Copilot notebooks do well

A Copilot notebook is a contained research environment. You bring the sources — a contract, a research report, a set of transcripts — and the notebook answers questions about them without drawing on anything outside that set. That containment is the point. When you need an answer grounded in specific documents, a notebook is more reliable than a general chat session because you control what it can read.

Notebooks work well for preparing a summary of a long document, comparing two reports, drafting a section of a proposal from source material, or answering a specific question before a meeting. They are not designed for recurring tasks, for acting on your calendar, or for handling requests that arrive from other people.

For a comparison with other source-grounded tools, Copilot notebooks vs. ChatGPT Projects covers how the two differ in source handling and instruction persistence.

What Copilot agents do well

A Copilot agent is configured to handle a defined task type across your Microsoft 365 environment. It has a system prompt that describes its role, a set of knowledge sources it can read, and tools — connectors to email, calendar, Planner, SharePoint, or other services — that let it take action. The agent runs when triggered, either by a user message, a scheduled job, or an event in a connected system.

Agents work well for triage workflows (routing incoming requests to the right queue), recurring tasks (sending a weekly summary, creating review tasks from a log), and cross-tool processes where the result in one system should cause an action in another. They require more careful design than a notebook because their scope is not naturally bounded by a set of documents.

Microsoft Copilot agents covers how to build and publish agents in Copilot Studio. GitHub Copilot agents vs. skills covers a related distinction in the developer tooling side.

When to use a notebook

• You need answers from a defined set of documents and do not want the AI to draw from general knowledge or other sources.
• You are preparing a draft, a summary, or a comparison that depends on specific source material.
• You want to share a curated reading environment with a colleague without giving access to broader Microsoft 365 data.
• The session is one-off research, not a recurring task.

When to use an agent

• The task involves creating, updating, or sending something in a Microsoft 365 service.
• The task recurs: daily triage, weekly summaries, routine routing.
• The trigger is an event in another system, not a person typing a question.
• Multiple people need to interact with the same configured assistant across different conversations.

Using notebooks and agents together

A notebook and an agent can complement each other in a workflow. Research done in a notebook — reading a contract, summarizing a long document, extracting key terms — produces output that can inform the instructions or knowledge sources you give an agent. The notebook is the preparation; the agent is the execution.

The gap is handoff. The notebook session does not automatically update the agent's context. You need to decide what information from the notebook is worth carrying forward, put it somewhere the agent can read it, and keep that record current. A shared project record in fenbs is one way to do that when the agent connects via a supported AI app.

How to do this with fenbs

1. Create a free account and add a project for the work this notebook or agent supports.
2. After a notebook session, add the key decisions and next steps to fenbs as Decisions and rules and tasks.
3. Before configuring an agent, write the scope, approved actions, and escalation rules as Decisions and rules in the project.
4. Connect a supported AI app — Claude, ChatGPT, or Cursor — to fenbs over MCP from Integrations, and ask it to read the project context before each session.
5. At the end of each session, save where you left off so the next session starts from the current state.