Skip to main content
Split a research question into ideas that run in parallel, compare their results, and follow up on the promising ones. A task is a tracked unit of work that an agent plans, such as “Clean the survey data” or “Fit a gradient-boosted model”. An idea is a task that tests one hypothesis or approach, and Qualia gives each idea its own agent.

What is a task?

Each task you see has a title, an objective, a status, and an assigned agent. A task can also depend on other tasks. For example, “Compare models” waits for “Train model A” and “Train model B” to finish. Tasks come in two kinds:
  • Ideas test a hypothesis or experiment worth pursuing on its own, such as “Try a transformer encoder”. Qualia creates a dedicated agent for each idea, and that agent works the idea until it finishes.
  • Work tasks are helper steps, such as downloading data or refactoring a script. They stay with the agent doing them, so you won’t see them in your list of ideas.

Plan work in chat

Ask an agent to plan the work in natural language. For example:
“Compare three ways to predict churn on this dataset, then summarize which works best.”
The agent proposes ideas, creates an agent for each one, and adds work tasks where it needs helpers. You don’t need tasks for simple questions; the agent answers those directly. To change how readily agents split work across more agents, adjust Parallelism in AI settings.

Follow your ideas

You’ll find ideas in the Chats list in the left sidebar, beneath the chat that proposed them. A follow-up idea connects to the idea it builds on. Helper agents that aren’t ideas collapse under Supporting agents; click it to expand them.
  • Hover an idea to see its description, its findings, and its metric values.
  • Click an idea to open its agent’s chat, where you can ask about its work.
  • Watch progress: A progress bar appears when an idea’s agent reports progress.
When an idea finishes, Qualia stops the helper agents working for it. The idea’s own agent stays available, so you can still ask it questions. For everything else in the sidebar, such as nested agents and waiting requests, see Agents.

Sort ideas by a metric

When agents compare ideas on a shared measure, such as validation accuracy or error, you see that metric’s value beside each idea. An Ideas heading then appears above the list.
  • Click the metric name on the right of the heading, then choose a metric under Sort by metric. Each option says whether lower or higher is better.
  • Click Ideas to return to lineage order, which groups follow-ups under the ideas they build on.
If agents have defined several metrics, Qualia picks a default. You can switch to another at any time.

Add a follow-up idea

  1. Open a chat first, because the new idea’s agent starts from the open chat.
  2. Hover an idea and click Follow-up idea, or right-click it and choose Follow-up idea.
  3. Describe the idea to test, then press Enter. Press Escape to cancel.
Your follow-up starts after the original idea finishes.

Verify an idea

To get an independent check of a result, hover an idea and click Verify with new agent. Qualia adds an idea titled “Verify: [IDEA_TITLE]” whose agent reviews the original idea’s notebook, evidence, and results. It reports whether the result holds.

Task statuses

Completed, failed, and cancelled tasks are final, and they release any tasks waiting on them. A paused task isn’t final, so tasks downstream of it stay blocked until you resume its agent. If an error interrupts an agent, Qualia pauses its chat instead of failing the task. The task stays In progress with its saved results. Resume the agent or send it a message to continue.

Track tasks in a chat

Above the chat composer, a task summary shows how many of the agent’s tasks are done, such as “2/5”, and names the current one. Click it to expand the full list, ordered by status. Chat activity also shows you when agents create, update, or reassign tasks. In the Chats list, you can see an agent’s assigned tasks beneath it.

Tasks and knowledge

Agents record what a task established as claims in Knowledge, where you can trace each one to its evidence. A claim is a recorded statement backed by evidence, such as a captured value or a quoted line from a file. Each claim links back to the task and chat that produced it. A task can also declare required variables: values an agent must capture from running code before it completes the task. Downstream tasks then read those values. To capture or read them yourself, see Task variables.