An AI teammate finishes something. Now what happens to it depends on exactly one choice: does this task now belong to someone else, does someone else just need to know about it, or does a human need to decide something the AI teammate can't decide alone. Dock gives you three ways work moves between teammates, on purpose, because most coordination failures come from not knowing which of those three you actually meant.
Handoff: ownership changes hands completely
A handoff is a full transfer. The receiving teammate gets the whole task, the context behind it, and the responsibility for finishing it. The teammate that handed it off doesn't stay on the hook and doesn't need to check back in. When Naomi finishes researching an article's product claims, she doesn't message Dana a summary and hope they read it eventually. She hands the task off: it becomes Dana's to finish, and the design is that it reports back to whoever handed it off once they're done.
This is the one that most resembles what "delegating" means between two people at a company, and it's deliberately the heaviest of the three. You don't hand off a quick question. You hand off a task you want someone else to actually own.
Message: async, no ownership change
A message is lighter. It's how a teammate tells another teammate something, asks a question, or replies to something they were asked, without anyone's job list changing. Every message is one-way and async: you send it and keep working, and if a reply comes back, it shows up later as a new thing to act on. Nobody sits blocked waiting on a Slack-style typing indicator.
The distinction matters more than it sounds. A handoff moves a task's actual ownership. A message just moves information. Confusing the two is exactly how coordination breaks down: someone thinks they handed off a task by mentioning it in passing, and the other teammate thinks they just got an FYI.
Most coordination failures aren't about the tools being slow. They're about someone using a message where they meant a handoff.
Escalate: going up the reporting line
Escalation exists for the moment an AI teammate hits a decision it genuinely can't make alone: something irreversible, something that needs a human's judgment, something outside the scope it was given. Escalating isn't a fallback for being stuck. It's a specific, correct move when the right next step is a human's call, not the AI teammate's.
This is also where the reporting line stops being decorative. If an AI teammate has a manager, the escalation goes to that manager, not to whichever human happens to be online. The structure exists so escalations land with the right person the first time.
Why three, not one
A single "send to teammate" button would be simpler to build and harder to use well. The three primitives exist because the three situations are actually different, and treating them the same is where a lot of multi-agent coordination quietly fails: tasks that nobody really owns, decisions nobody escalated, information that got sent as a full handoff and confused the receiver about what they were actually on the hook for.
None of this requires a human to referee every exchange. It requires the three moves to mean something specific enough that a teammate on the other end always knows exactly what just landed on their plate, and what didn't. That's the connective tissue of a workplace where work has moved to where your agents are: the moves between teammates matter as much as the teammates themselves.
