PA or specialist: choosing between Copilot Cowork and Agents

Recently I had the opportunity to play with Copilot Cowork. And during the last Power Addicts we got an epic Copilot Cowork demo by Daniel Laskewitz. However during the break some of us had a discussion the never ending question on when to use what. Something I almost always answer with: well, it depends.

So I like to use a simple metaphor to pull them apart:

• A Coworker (Copilot Cowork) is your personal assistant (or Virtual Assistant for that matter).
• An Agent (Copilot Agents) is a colleague you’ve trained up on one specific (set of) task(s).

Both are equally useful. Both are AI. But you bring them in for very different things, and yes you can use Copilot Cowork for everything and burn tokens.. But just because you can, doesn’t mean you should!

The PA: Copilot Cowork

Picture it like this: you walk over to your assistant and say “quickly draft an email to that client” or “find me the latest sales report and combine it with market trends”. No preparation needed, no briefing document up front. You ask your question, and you get help right away.

That’s exactly what a Coworker does. It works well for:

1. One-off questions
2. Unpredictable situations
3. Exploring and creating

The nice thing is you don’t have to set anything up. No knowledge base, no connectors, no rules. It thinks along with you in the moment. The flip side, yes there is one, is that it doesn’t take on a structural role. Ask the same thing tomorrow and you start from scratch again.

At the time of writing, you can extend a Virtual Assistant with skills, but there is a cap of 20 skills (for now), and those skills stay personal to you. They are not shared automatically with others, while with Agents you can share them across a team. So if you have a skill that is useful for others, you can share it manually, but they just get a copy. That also means that Coworkers are really personal assistants, while Agents are more like colleagues.

Example scenario: Brainstorming a campaign, or researching something you haven’t explored before. Knowledge work that’s a little bit different every time.

The specialist: Copilot Agents

An Agent is a different story. This isn’t an assistant you steer question by question, but a team member you’ve trained up in advance on exactly one task. And just like with a new colleague, that onboarding takes time: you give it knowledge, you connect it to your systems, you hand it the ground rules and structure.

But after that, it can act on its own within that defined scope. It will act somewhat consistently depending on your rules. An Agent works well for:

1. Repeatable processes
2. Clear business flows
3. Tasks that involve specific knowledge or integrations

That is also why Agents fit team operations better: once configured, they can be shared with a team to produce repeatable, consistent outcomes inside a defined process.

Example scenario: Automatically categorising emails, or writing a meeting summary in a strict format you already decided on before, an onboarding assistant for HR, a content pipeline for marketing. All things you do more than once, in a way you can reasonably describe.

And that’s where the heart of the story lies.

Predictability versus variation

If I have to boil it down to a single measure, it’s this one: how predictable is what you’re asking for? The more a task looks like a fixed workflow, the more you move towards an Agent. The more unique and changeable the question, the better a Coworker fits.

A handy rule of thumb I use myself: if you can describe it as a process, it’s probably an Agent.

From chat to systems thinking

I think of it as two layers:

• Coworkers are your productivity layer: they make individual people faster.
• Agents are your automation layer: they take over parts of a process.

The key insight: an Agent isn’t a “better chat”, it is an architecture choice. And the moment you see it that way, you start making different decisions about governance, maintenance and ownership. Who trains this Agent? Who keeps it current? What happens when the process changes? Ultimately, however you slice it, someone is responsible for what that agent does. Don’t underestimate that, and don’t forget that you can still build workflows.

So how do you choose?

No complicated model needed. Two short checklists.

Reach for a Copilot Cowork when:

• your question is new or unique
• the context is different every time
• you want to explore or create

Reach for an Copilot Agent when:

• you do the same thing five times or more
• the input and output are reasonably predictable
• there’s value in consistency

Stuck between the two? That hesitation is often the answer in itself: start with a Coworker, learn what the process really is, and only build an Agent once the pattern repeats. It varies per scenario, but as a guideline it serves you well.

To close: it’s not either/or

The question, in the end, isn’t “Cowork or Agents?” It’s “when which?” The organisations that win at this do both: they use Copilot Coworkers for knowledge work that’s different every time, and Copilot Agents for the scale and automation of what’s repeatable.

And maybe that’s the real shift. AI gets genuinely interesting the moment you stop thinking in prompts and start thinking in roles. Who actually does what here and can I leave that to an assistant, or does it call for a trained-up specialist? Once you look at your work that way, the choice between Copilot Cowork and Copilot Agents suddenly becomes a good deal less complicated.

So in the end, it still depends 🤷