Using Agents

There are two ways to work with a server-side agent: mention it in a thread, or chat with it in the AI Assistant panel.

Mention an agent in a thread

Anywhere you can write a message — a task's Discussion tab, a brainstorm, or a channel — type @, pick the agent, and write your request:

@plan-bot break this task into subtasks and create them

What you'll see:

  1. Your message posts immediately, and the thread shows a "thinking" indicator.
  2. The agent works — it may make several provider calls and call tools (read the task, search the knowledge base, create or update entities).
  3. The agent posts its final reply as a normal message, and the thinking indicator clears.

Anything the agent changes (tasks created, fields updated, pages written) happens for real and shows up in the relevant activity feed, attributed to you.

Good things to ask an agent in context:

  • On a task: "summarize this thread", "draft a subtask breakdown", "what does this code do and why".
  • In a brainstorm: "turn this idea into a project and tasks".
  • In a channel: "what's the status of the migration project?"

When a request matches one of the workspace's AI Workflows, the agent follows that process rather than improvising — ask it to "create a task for X" and the task comes out per your task_create conventions: the right team, labels, estimate, assignee, priority, and acceptance criteria. See "AI Workflows".

The AI Assistant panel

The sparkle icon in the bottom bar opens the AI Assistant — a quick, one-on-one chat with a workspace assistant, separate from any task or channel. Use it for quick questions and analysis.

  • Pick the assistant — choose among the agents marked Available as workspace assistant (see "Configuring an Agent").
  • Threads — start a new chat or return to a previous one; the panel keeps your history.
  • Dock or float — pin the panel to the side or let it float over the app.

The assistant is the same kind of agent with the same access — it can read and change workspace data through tools, within your permissions.

If a run fails

If an agent's provider key is missing or invalid, or the model is unavailable, the run fails and says so. Check the agent's configuration in Settings → Agents (key, provider, model). The "max iterations" cap can also stop a long run; raise it if a legitimately multi-step task needs more room.