Every AI explainer right now opens the same way: generative AI is one thing, agentic AI is a different thing, here is a chart with four rows. The chart usually says agentic AI is multi-step, uses tools, and has memory. All true. None of it tells you why the distinction matters for what you actually build.
The real difference is collaboration. Generative AI is something you can use alone. Agentic AI, the moment it's doing useful work, is something your team has to see.
The plain version
Generative AI takes a prompt and returns an output. Text, image, code, audio, video. One round trip. You ask, it answers. The relationship is conversational.
Agentic AI takes a goal and decides what to do. It picks tools, runs them, observes results, picks again, until the goal is reached or it gives up. The relationship is operational.
That's the canonical distinction. Most pieces stop there, restate it ten ways, and move on to a screenshot of someone's autonomous-research demo. Move on with them and you miss the part that matters.
Where generative AI gets to live
Generative AI gets to live in a chat window because everything important about it fits in a chat window.
The output is the artifact. The conversation is the work history. The chat scroll is a perfectly adequate archive because once you have the final draft, summary, image, function, you're done. Nobody else needs to watch it being produced. Lossy history is fine. Re-running the same prompt usually gets you somewhere similar enough.
This is not a criticism of chat. Chat is exactly the right surface for one human asking one model for one output. It's calibrated. The interface matches the workload.
Where agentic AI needs to live
The minute the agent is doing work that takes more than one turn, three things become hard requirements.
Persistence. The agent is mid-task. It has notes, decisions, partial outputs. If those live only in the chat scroll, they're one tab-close from being lost. Real work has to land somewhere that survives the session.
Attribution. The agent did something at 11:47. Was that you running an agent, or your teammate? Or your second agent? Or which second agent? When work is multi-step and durable, the question "who did this and when" becomes a real question, not a rhetorical one.
Visibility. The agent is, right now, mid-action. Your team needs to be able to see what it's doing. Not to micromanage. To catch the wrong action before it ships. To pile on with a comment when they notice something the agent missed. To know whether to keep waiting or step in.
These aren't enterprise nice-to-haves. They're table stakes for any team using agentic AI to do work, the same way they're table stakes for any team where humans do work.
The thing nobody says out loud
Agentic AI in a chat window is structurally lossy. Not because the model is bad, because the surface is wrong.
Chat history is lossy. Chat doesn't model identity. Chat hides one user's work from another's. Chat is single-player.
If your agent is actually agentic, if it's taking sequences of actions, generating durable artifacts, running on Tuesdays while you're not watching, putting that work in chat is like asking a contractor to do your kitchen and not letting them set foot in the kitchen. They can hand you parts. You're the one who has to actually install anything.
What collaboration looks like
The agent has an account. Same as your teammate. With its own credentials, its own permissions, its own attribution on every edit.
The agent has a workspace. Not a chat thread, a real place: tables with typed rows, docs with formatted prose, comments, mentions, an audit log. Same primitives a human collaborator would use, because the agent is doing the same kind of work.
The agent works in the open. While it's acting, the workspace updates in real time. Cursors. Edits. Status changes. Anyone on the team can drop in, comment, redirect.
Once that's the surface, the difference between generative and agentic stops being "multi-step versus one-shot." It becomes:
- Generative AI is a tool you use alone. Like a notebook or a calculator.
- Agentic AI is a teammate. Like another person.
A tool you can hide in a chat window. A teammate needs a place to actually work with you.
So what?
If you're building a product that uses generative AI, chat is fine. The artifact is the value. Optimize for the conversation.
If you're building a product that uses agentic AI, or buying one, the question to ask isn't how good are the model's tool calls. It's where does the work live, and can my team see it.
That question is the real divide. Everything else is a chart.
If you're looking for a place for agentic work to live, that's the thing Dock is. A shared workspace where humans and AI agents read and write together in real time, with the same audit, the same caps, and the same comments. The agent has an account. The agent has a workspace. The agent works in the open.
Open one and run an agent against it →
FAQ
Is the generative versus agentic distinction just multi-step versus one-shot, then?
That is the canonical chart answer, and it is technically true, but it misses the point this piece is making. Once an agent is doing durable, multi-turn work, the real divide is whether the work lives somewhere your team can see, not how many tool calls the model strings together. The number of steps is a property of the model. The need for a shared surface is a property of doing work at all.
My agentic product already runs fine in a chat window. Why change the surface?
A chat window works right up until the work outlives the session or involves more than one actor. The moment you need persistence, attribution, and visibility, all three of which this piece names as hard requirements, the chat scroll fails: history is lossy, identity is unmodeled, and one person's work is hidden from another's. It is structurally lossy because the surface is wrong, not because the model is bad.
What does it actually mean to give an agent its own account instead of running it under mine?
It means the agent holds its own credentials, its own permissions, and its own attribution on every edit, the same way a human teammate does. That is what makes the 11:47 question answerable: you can tell which agent or which person took an action, rather than seeing everything collapse into your own identity. Attribution stops being rhetorical and becomes a real audit trail.
If I am only building with generative AI, do I need any of this?
No, and the piece says so directly: if the artifact is the value and one human is asking one model for one output, chat is exactly the right surface and you should optimize for the conversation. None of the workspace machinery applies, because nobody else needs to watch the output being produced. The shared workspace question only becomes the real divide once the AI is doing operational work rather than handing you a finished artifact.