Skip to main content
Run several experiments at once by delegating each to its own agent. This page covers how to start agents, run them in parallel, and manage them from the Chats list in the left sidebar.

What is an agent?

An agent is a chat that works on its own. Each agent has its own conversation history and can attach to a notebook to write and run code. You can assign it tasks, the tracked units of work agents plan. Every agent appears in the Chats section of the left sidebar.

Start an agent

To start an agent, type a message in an empty chat. You can also click New agent (+) in the Chats section header, or right-click blank space in the list. In a shared workspace, creating an agent requires edit access. Your agents also create other agents when work splits well, as described in Parallel agents.

Execution mode

Each agent runs in one of two execution modes:
  • Notebook runs code in an attached notebook, so you see cells and kernel-backed results. A kernel is the process that runs the notebook’s code.
  • Script creates and runs standalone files and shell commands. It suits training runs, scraping, batch jobs, and other work that doesn’t need cells.
Set the default under Default agent execution mode in Settings > Agents. An agent keeps the mode it started with, so if you change the setting, existing chats don’t change. Chats started from a notebook always use Notebook mode.

Parallel agents

Sequential experiments are slow: you try one thing, wait, check the result, then try the next. Instead, ask Qualia to run the variants side by side. For example:
“Compare three approaches: (1) random forest with default hyperparameters, (2) XGBoost with early stopping, and (3) a small neural network. Use 5-fold cross-validation for each. Run them in parallel.”
Qualia proposes three ideas, one per approach, and gives each its own agent. An idea is a task that tests one hypothesis; see Ideas and tasks. When they finish, you can compare the results side by side. The same pattern works for a hyperparameter sweep, such as three learning rates, or for comparing feature sets. The Parallelism control sets how readily Qualia splits work across agents when you don’t ask it to. Change it per chat from Research style in the chat box, or set a default in Settings > Agents. An agent created by another agent inherits its execution mode, verbosity, and auto-approve setting. If the parent runs in autonomous mode, the new agent does too. Its model comes from the Sub-agent model setting; see Agent models.

Approve new agents

To review each new agent before it starts, turn on New agents require user confirmation in Settings > Agents. The creating agent then shows an approval card in its chat, and the new agent starts only after you approve. This setting is off by default.

The Chats sidebar

The Chats section of the left sidebar lists your agents and the ideas they’re working on. You see agents started by another agent nested under it. Click the arrow, or press Right and Left, to expand or collapse a branch.
  • Requests: A bell marks an agent that needs input or approval. Hover over it to see its full question or command. The chat’s creator can accept or reject command approvals.
  • Tasks: Assigned tasks appear beneath the agent. See Ideas and tasks for how ideas appear and sort.
Agents that need your input move to the top of the list. While you work in one chat, a reminder tells you when other agents are waiting. Use it to open the next waiting request. Pipeline runs appear under Pipelines, grouped by pipeline. A pipeline is a saved set of instructions that produces a report or other artifact each time it runs; see Pipelines. Expand a group to open a run’s conversation. Use Load more agents to reach older runs and chats. Use Tab and Shift+Tab to move between agent cards and their task previews, and Enter or Space to open a focused agent. If the list shows Couldn’t load agents, wait a moment and click Retry. If loading more agents fails, your loaded agents stay visible; click Retry loading agents below them.

Agent actions

  • Open: Click an agent to open its conversation.
  • Rename or export: Right-click an agent and choose Rename agent or Export chat.
  • Slack updates: Right-click an agent and choose Notify in Slack, or Connect Slack… if Slack isn’t connected yet. See Slack.
  • Delete: Right-click an agent and choose Delete agent to remove it, its history, and any attachments no other chat uses.
  • Select several: Cmd-click or Ctrl-click, Shift-click, or right-click and choose Select agent. Check the agents you want, then click Delete. Press Escape to leave selection mode.
To stop or resume one agent, open its chat and use Stop generation or Resume agent in the chat input footer. To stop or resume every agent in the workspace, open the Agents menu and choose Stop All Agents… or Resume All Agents…. In Qualia Cloud, a workspace owner stops all running agents, and other members stop their own. Resume all applies to your own chats. If an account’s usage limit or payment requirements prevent an agent from running, the agent pauses with its queued messages and tasks preserved. Automatic recovery does not retry the agent or replace it with another agent. Once the account can run again, resume the agent or send a new message to continue.

Drag agents into tabs

Agent cards work like tab handles. Drag a card out of the sidebar to place that agent’s conversation where you want it:
  • Drop it onto a tab bar to open the conversation as a tab there. Drop between tabs to choose its position.
  • Drop it near the edge of the editor area to open it in a new split.
  • Drop it onto a chat box to mention that agent in your message instead.
If the conversation is already open, dragging its card moves the existing tab to where you drop it.

Agent communication

Agents can send each other messages, so one agent can hand results or questions to another. Click a link to an agent in a chat to open that agent. All agents in a workspace share the same knowledge, the claims and evidence Qualia records as it works. Open it with the Knowledge button (lightbulb) in the titlebar, or View > Knowledge.

Agents and tasks

Each task has an assigned agent, and each agent works through its own queue. Every idea gets a dedicated agent that works it until it finishes. See Ideas and tasks for how agents plan and track work.

Autonomous work

To let agents work for hours without stopping for approval, use autonomous mode. Agents in autonomous mode can post progress to Slack.