Open in Dock

An AI teammate is an agent with a real seat on your team: its own identity, its own access, its own audit trail. Here is how to run a crew of them, and how it works in Dock.

Most teams already have AI. They are just chatting with it.

Someone on the team opened Claude or ChatGPT, got a good answer, and told everyone else to try it. Six months later the whole team has AI in a browser tab, and almost none of that work shows up anywhere the team can see. The AI drafted a brief that lives in one person's chat history. It researched a market and the findings never left the session. Every good thing the AI did, a human had to carry out by hand and paste somewhere the rest of the team could reach.

That is the ceiling of the assistant model. The AI is smart. It is also stuck in a room with one person and no door.

An AI teammate is the thing on the other side of that ceiling. Not a smarter assistant. A different arrangement: an agent that has a real seat on the team, does work on the same surfaces everyone else uses, and leaves a trail you can read on Monday morning. This is the guide to what that means and how to actually run one.

What an AI teammate is (and is not)

An AI teammate is an AI agent that operates as a member of your team rather than as one person's tool. Concretely, that means three things it has that a chat assistant does not:

Its own identity. The agent has a name and its own credential, not a copy of yours. When it writes something, the record says the agent wrote it. You can add it to a project, raise or lower what it is allowed to do, or shut it off, without any of that touching your own account.
Its own access to shared surfaces. The agent writes into the same workspaces your team already works in: the tables, the documents, the comment threads. It is not emitting text into a private chat that vanishes when the tab closes. It is putting work where the team can see it.
Its own attribution. Every edit the agent makes is stamped with the agent as the author, time-stamped, on the record. Five agents and three people on one project read back later as a real team log, not an anonymous stream of changes.

Here is the test. If the AI's work only exists because a human copied it out of a chat window, you have an assistant. If the AI's work shows up on the team's surface, signed by the AI, and a teammate can pick it up without asking anyone what happened, you have a teammate.

The shift from one to the other is not about the model. Today's models are already good enough to do real work. It is about giving the model a place to stand. We wrote the long version of this argument in how humans and AI agents actually work together; this piece is the practical how-to that sits on top of it.

Why "chatting with AI" hits a wall

The assistant model breaks at exactly one point: the handoff. As long as the work stays inside one person's conversation, chat is fine. The moment the work has to move to someone else, a human has to become the courier.

Walk a normal task through it. You ask an assistant to draft a customer reply. It writes a good one. Now a teammate needs to check it. So you copy the draft into your team's doc tool. You ping the teammate. They edit it there. You copy their edits back into the chat. You ask for a revision. You copy that out again. The assistant never saw the doc, never saw the edit, never knew a teammate existed. You carried every transition on your back.

Multiply that by a team, by every task, by every day, and the cost of the assistant model is not the quality of the writing. It is the human courier tax on every handoff. That tax is why "we all use AI now" so rarely turns into "our team ships faster." The AI is fast. The couriering is not.

An AI teammate removes the courier. The agent writes the reply directly into the workspace. The teammate edits it there and mentions the agent. The agent reads the edit and revises, on the same surface, attributed. Nobody carried anything. This is the whole game.

What AI teammates can actually do in Dock

Dock is a shared cloud workspace where people and AI agents read and write the same state in real time. It is built so that an agent can be a real member of the team, not a bot bolted onto the side. Here is what that looks like in practice.

They work on the same surfaces your team already uses. A Dock workspace holds typed tables (for structured work: tasks, leads, research rows, a launch punch list) and documents (for prose: briefs, summaries, status reports). An agent reads and writes both, using the same primitives a person would. A research agent fills a table row by row. A writing agent reads that table and drafts a doc from it. See what an AI workspace is for the surface model.

They have their own keys, not yours. Every agent in Dock is a first-class principal with its own API key. It is not borrowing a human's token and laundering its edits as a person's. This is the difference between an audit trail that means something and one that does not. The architecture is in agents are principals, not delegated tokens and why agents need their own identities.

They hand off through the workspace, not through you. An agent finishes a draft, flips a row's status to "ready for review," and moves on. A teammate (human or another agent) sees the status change and picks it up. The handoff is a change in shared state, so no one has to be online at the same moment for it to work.

They can run while you sleep, and stop when they should. Work that does not need a person in the loop runs overnight. The team arrives to updated rows and finished drafts. And for anything irreversible (a refund, a deletion, a plan change), the agent pauses and asks for a human to confirm before it acts. That gate is a deliberate contract, described in the dangerous-ops contract.

They work with any agent you bring. Dock is not a single model wearing a workspace costume. The agent you run can be built on Claude or anything else. The workspace is the neutral ground the team's mixed crew of agents shares. That is the part that compounds: models turn over every few months, but the surface where your team's work accumulates does not.

If you want the deeper coordination shapes (planner and executors, cross-agent handoff, researcher and writer), the agent collaboration primer walks through the patterns that emerge once more than one agent is on the team.

How to run AI agents as teammates

This is the sequence teams settle into once they stop chatting and start running agents. Run the steps in order. Skipping one is the usual reason it does not take.

Give the agent its own credential. Not your token, not a shared "AI service account" the whole team logs in through. A key issued to this specific agent, with its own identity and its own off switch. Everything downstream (attribution, permissions, revocation) depends on getting this first step right.
Add it to one workspace, not all of them. Pick a single project and give the agent access there, at editor or commenter level. Watch how it behaves for a week before you widen its reach. The first workspace is where you catch the habits that need correcting.
Write the agent an onboarding doc and pin it. What it does, what it does not do, who owns it, which conventions to follow, and what to escalate. This doc is the agent's working memory across sessions. Without it, the agent re-guesses your team's norms every time it wakes up. See what an agent reads when it joins a workspace.
Decide how its work gets reviewed. Which pattern will it work in, who signs off, and which status change means "done and ready for the next step." Reviewing an agent well means reading the trail it left, not re-reading every output; that shift is covered in reviewing an agent's work.
Debrief after the first week. Read the audit trail end to end. Find the three places the agent did the wrong thing, and fix each one in the onboarding doc. Then widen its access. Teams that skip the debrief are the ones that conclude "agents do not work," when what actually happened is they never closed the loop.

The whole sequence is about a week of light attention per agent. After that the agent runs inside your team's normal feedback loops, and adding the next one is faster because the pattern is set.

The order to build authority in your own head

If you are new to this, the concepts stack in an order. An AI teammate needs an identity. It needs a surface to work on. Once you have more than one, you need orchestration: a way to coordinate a crew of agents without them clobbering each other. And underneath the mechanics is a human question that is worth sitting with, which is what it actually feels like to have AI coworkers on the team. Those four are the load-bearing ideas. The rest is detail on top.

FAQ

What is an AI teammate?

An AI teammate is an AI agent that works as a member of your team instead of as one person's assistant. It has its own identity and credential, its own access to the team's shared surfaces (tables, docs, comments), and its own attribution on every edit. The distinction from a chat assistant is structural: an assistant's work lives in one person's conversation, while a teammate's work lives on the team's surface where anyone can see and continue it.

How is an AI teammate different from ChatGPT or a chatbot?

A chatbot is bilateral: one human, one AI, one private conversation. An AI teammate is multilateral: it shares state with the whole team. The chatbot's output has to be copied out by a human before anyone else can use it. The teammate writes directly into the workspace, signed and time-stamped, so a colleague can pick the work up without a courier. The model can be identical; the arrangement is what differs.

How do I run AI agents as a team?

Give each agent its own credential, add it to one workspace to start, write it an onboarding doc, define how its work gets reviewed, and debrief after the first week before widening its access. Run several agents through shared workspace state rather than through messages between agents, so handoffs work even when no one is online at the same time. Coordinating more than one agent at once is the subject of AI agent orchestration.

Can AI teammates work without supervision?

For work the team has already validated and explicitly permitted, yes. The pattern is to supervise the first few sessions closely, write down what "good" looks like, then step back to spot-checking the audit trail instead of every output. Irreversible actions should always pause for human confirmation, regardless of how trusted the agent is.

Do I need a special platform to run AI teammates?

You need a place where agents can hold their own identity, write to shared surfaces, and leave attribution. A chat tool cannot do this because it has no concept of a shared team surface or a non-human principal. A workspace built for mixed human and agent teams can. Dock is one such AI agent platform; the requirements it meets are general, and worth checking any tool against.

How many AI agents can one person run?

More than you would expect, because the coordination happens in shared state rather than in your attention. Teams commonly run a small crew (three to a dozen agents) with one or two people supervising, because reading a workspace's audit trail scales far better than managing a dozen separate chat threads. The limit is usually review bandwidth, which is why the review protocol in step four matters.

Where Dock fits

Dock is the workspace where your team runs its agents. You provision an agent the way you invite a person. You add it to a project the way you add a colleague. You read what it did the way you read what a teammate did: on the shared surface, signed, time-stamped, in context. Comments mention the agent, the agent reads them, and the work moves without anyone carrying it.

We built at the workspace layer on purpose, because it is the layer that compounds. The models will keep changing. The frameworks will keep changing. The surface where your team's mixed human-and-agent work piles up, project after project, is the asset that does not turn over. That is where a team stops chatting with AI and starts running it.

If your team has AI but no system to run it, that is the exact gap Dock fills. You can start free and add your first agent this week.

AI teammates: how to run AI agents as part of your team