AI is on the team now. The shape of work has to change.

Most products bolting AI onto their app picked the easiest metaphor: the chat thread. You message the agent. The agent messages back. It's a tidy fiction that lets shippable demos happen in a weekend.

But chat is a one-to-one channel. It's optimized for asking and answering, not for shared state. When your agent edits a doc, the next human reading it can't see the edit history through the chat. When the agent closes a ticket, the next agent picking up adjacent work has no memory of the decision. Each session starts cold.

Chat doesn't scale to teams of agents because chat was never a collaboration substrate. It's a transport.

Look at how human teams collaborate. Not over chat — over shared state. A roadmap doc both editors can rewrite. A tracker both PMs can update. A spec both engineers refine. People talk about the work, but the work itself lives in the artifact, persistent, addressable, with history.

Agents need the same thing. Not a transcript of what they said, but a workspace they can read, write, and hand off to the next human or agent. The workspace is the substrate. Conversation is optional.

That's the bet behind Dock: the next decade of software gets built when humans and agents share workspace state, not when they trade messages.

Every shipped agent today is a forgery in disguise. It has its user's API key, its user's permissions, its user's blast radius. When it acts, the audit log says the user did it. When it screws up, the user owns the bill. When it's compromised, it's the user's whole stack at risk.

We made agents borrow human credentials because we didn't know what an agent was. They were tools, scripts, helpers. Things you ran. Borrowing keys was fine when the agent took twenty actions per session.

It's not fine when the agent is taking thousands. The agent needs its own identity, its own scopes, its own audit trail, and an owner accountable for it. Same way an employee has a badge, a team, a manager. Without that, you can't run a fleet. You can run one helper. That's the ceiling.

One human + one assistant is the demo. The product is one human + a team of agents, working in parallel, on real shared state, with humans reviewing and intervening where it matters.

That changes everything about the surface. The workspace has to be multi-author by default. Identity has to be first-class for non-humans. Authorization has to gate dangerous ops with a consent loop, not a yolo prompt. The version history has to tell you which actor — human or agent — touched which line.

The product is not "AI inside your existing tools." It's a new tool, designed from the inside out for plural collaboration across kinds.

Dock is the AI workspace for you, your team, and every agent you run. Tables and docs that read + write the same state. Agents as first-class identities with their own keys and scopes. Two-key handshakes on dangerous operations. Signed agent inheritance so accountability never gets diluted.

We think this is the right shape. The blog at trydock.ai/blog is where we're showing our work — the field notes, the tradeoffs, the engineering decisions. Read what's there. Tell us where we're wrong.