I started calling it gangprompting because “shared agent session” makes the idea sound more formal than it is. One person opens a thread, starts exploring a problem, and gets the agent moving. Then someone else who knows the domain joins the same conversation and points it in a better direction.

That second person is not receiving a summary in a new chat. They are entering the work where it already exists: the initial question, the assumptions, the agent’s responses, and the places where the conversation started drifting. The context is shared, so the handoff can happen inside the problem instead of around it.

Starting before I have the expertise

The clearest example began with an idea I did not know much about. I could still open a thread and establish the shape of the question. That early pass was useful because it turned a vague idea into something another person could inspect.

Then Dax joined with knowledge I did not have. He could see what I had asked, what the agent had inferred, and where a more informed prompt would help. He guided the same agent rather than explaining the whole topic to me first and waiting for me to translate his explanation into another prompt.

That is the practical value for me: starting and steering become separate jobs. Curiosity is enough to begin. Expertise can enter at the moment it becomes necessary.

Shared context changes the handoff

Most collaboration around AI still happens through copy and paste. Someone sends a response in Slack, another person comments on it, and the first person returns to the agent with a compressed version of that feedback. Each hop loses detail. The person relaying the message also becomes an accidental interpreter.

Gangprompting shortens that loop. A knowledgeable collaborator can read the original exchange and speak directly into it. They can correct an assumption in its original context instead of correcting my recollection of it.

It also makes the division of labor visible. I may contribute the initial goal and product context. Someone else may contribute the domain constraint that changes the approach. The agent may help organize the exploration. Those are different contributions, and keeping them in one thread makes the sequence easier to follow.

I used the pattern twice within a few days because it fit the way the work was already happening. I do not expect every extra person to improve a session. A collaborator without a clear role can add noise just as easily as expertise. The useful moment is when someone can name a constraint or correct an assumption that the current conversation is missing.

A lightweight protocol for gangprompting

The pattern works best when the first person leaves the thread understandable. I try to state the actual goal, say what I do not know, and avoid presenting an early agent response as settled. That gives the next person room to challenge the direction rather than merely polish it.

When the knowledgeable person joins, the useful move is not “do this instead” without explanation. It is to add the missing constraint or point out the flawed assumption. That gives both the people and the agent a reason for the turn.

Conflicting instructions still need a human decision. A shared thread does not create a shared mind, and the agent should not be expected to decide which person has authority. If two directions diverge, the people involved need to resolve the goal and then state the decision clearly in the session.

Now, when I wander into an area I do not understand, I do not think of the session as mine just because I clicked “new session.” I can open the door, someone else can bring the missing knowledge, and we can both see how the work arrived at its current state. That is gangprompting.