Multi-chat

What multi-chat is

Multi-chat lets you split the chat panel into multiple panes, each connected to a different agent. Instead of one agent doing everything sequentially, you can have two or three agents working in parallel.

Splitting panes

Click the Split button at the top of the chat panel (or press Cmd+Shift+D). The panel divides in half, each side with its own chat thread and agent. You can split up to 4 panes.

Each pane shows the agent's role badge at the top, so you always know who you're talking to.

Shared context

All agents in a multi-chat session share:

• The filesystem. They see the same files on the pod
• The blackboard. A shared scratchpad where agents post summaries of what they've done
• Experiment data. All agents can read the same experiment logs

They don't share chat history. Each pane is a separate conversation. This is intentional. It keeps each agent focused on its task without noise from the others.

The blackboard

The blackboard is the coordination mechanism. When an agent finishes a task, it posts a summary to the blackboard. Other agents see it and can reference it.

For example: a Data agent might post "Finished preprocessing. 45,231 examples in /data/processed/train.jsonl". The Coder agent in the next pane sees this and knows where to find the data.

You can see the blackboard by clicking the Blackboard tab in the center panel.

When to use multiple agents

Parallel tasks. One agent preprocesses data while another writes the training loop. Neither is blocking the other.

Specialized work. A Reviewer agent in one pane reviewing code while a Coder agent in another pane continues building. The reviewer posts findings to the blackboard.

Monitoring. An Ops agent watching GPU metrics and system health in one pane while a Coder agent runs experiments in another.

When not to use it: If the work is inherently sequential (step 2 depends on step 1), multi-chat adds overhead. Stick to one agent.