Assistant, agent, agentic workflow — three words people muddle

Agents at Work — CC BY 4.0

Before I hand any work to an agent, I get clear on what I’m handing it to. Three words get thrown about as if they’re one thing — assistant, agent, agentic workflow. They aren’t, and the gap between them is exactly the gap in risk. This is Anchor 1 — continuous learning — in its most basic form: you can’t run a tool well until you know what kind of tool it is.

The three things

An assistant answers you. You open a chat, ask, read the reply, and decide what to do with it. It never touches anything unless you copy it out. You are in the room for every step. That’s the tool the Working with Claude course is about, and for a lot of work it’s all you need.

An agent acts for you. You set it up once — a task, the tools it’s allowed to use, the accounts it can reach — and it carries the work out itself: reads the inbox and drafts replies, pulls the figures and reconciles them, watches for a thing and does something when it happens. The shift is that it takes actions in the world, and often does it while you’re not watching.

An agentic workflow is several of those stitched together — one agent hands to the next, or one agent runs a chain of steps end to end. A quote comes in, gets read, priced, drafted, and queued for your approval, all without you touching each stage.

The move from the first to the third is a move along one line: how much acts on your behalf, unwatched. That’s the line this whole course runs along.

The distinction that actually matters

Here’s the one to hold onto, because it doesn’t change no matter how good these things get: capability is not agency.

A capable tool does impressive work. An agent, in the sense that matters to a business, would be a party that can decide, be answerable, and be held to account — hold authority, make representations, carry liability. Software isn’t that, anywhere. It can’t sign a contract in its own name, can’t be a company director, can’t be sued, can’t “stand behind” a decision. Capability is about what a thing can produce. Agency is about who can be held responsible. No amount of the first ever adds up to the second.

So when a tool drafts your quote or screens your applicants, there was never a second party in the room. You issued it; it’s yours. That sounds like a small philosophical point. It’s the practical foundation of everything here — and it’s Anchor 3, AI goodness at the foundation, in one line: you answer for what the agent does, so the job is to build it so you always can.

Why the words will keep sliding

There’s a catch that the fast pace makes real. The line between “assistant work” and “agent work” keeps moving up. The length and complexity of task these tools can carry out on their own has been climbing quickly — work that was firmly “you’d better do that yourself” a year ago is now something an agent will have a decent go at. So the categories aren’t fixed shelves; they’re a slope, and the tide is coming up it.

That’s not a reason to wait for the water to settle — it won’t, not for a while. It’s a reason to know exactly which rung you’re on right now, so you can step up deliberately rather than get carried. A task you’d happily let an assistant draft is not automatically one you’d let an agent send. The extra word — send, post, pay, reject — is where the risk lives.

What this means for the rest of the course

We’ll use “agent” as the plain word for “a tool that acts for you,” and we’ll be careful about how far it acts. Tier 1 is about deciding what’s fair game. Tier 2 designs it. Tier 3 builds and guards it. Tier 4 puts it to work and keeps you answerable for it.

Keep the one line running underneath: the cleverer it gets, the more it can do for you — and the more it matters that you stay the one who directs it, checks it, and owns the result. That’s not caution for its own sake. It’s what lets you hand over real work and sleep at night.