Use it

Tasks

One-off work by a team, as a chat with follow-ups, files and a live agent graph.

A task is one goal handed to a team. It opens as a chat. You see your messages, the team's reports, and the files it wrote (Word, PDF, Excel, slides, CSV, Markdown), each previewed in place.

  • Before: the cost preview estimates the team and its spend, so you can set a budget.
  • During: switch to Agents for the live graph, with spawns, messages, tool calls and statuses. You can switch the task's model mid-run and the whole team follows.
  • After: ask a follow-up and the root picks up with its whole history. Continue a task that stopped partway with more budget. Replay or fork it from its step journal.

Ask from any page. Every other page has tasks in a side panel: Ask in the sidebar, or Ctrl/⌘ K, shows and hides it. Ask how to write a script, what a setting does or anything else, right where you're working; follow-ups, files and the model picker work as here, and the conversation stays in the panel as you move between pages. Drag its edge to resize it, or float it over the page's corner from its header. Opened on a solution, a workflow or a task, it offers that as what you're asking about.

Ask about your data. A task can look up your organization's own data, read-only: "what's the latest record in @solution:helpdesk?", "how did the last run of @workflow:… go?", "why did @task:… fail?", "what changed between @task:… and its replay @task:…?", "what did we spend this month, by model?", "which customers used the assistant most?". Type @ to pick a solution, workflow or recent task. You see what any member sees, except private workflows: a task's conversation is visible to your whole organization, so they're never read into one. Your app's end users never get this: a customer can't ask your assistant for your analytics.

Which model answered. A conversation can switch models between follow-ups, so every answer shows the model that gave it. Hover it to see every model the team used for that answer. A workflow's chat does the same for each run's result, and the builder's Design with AI for each reply.

Images

Attach a screenshot, photo or chart with your initial request or a follow-up. A vision-capable OpenAI, Anthropic, Gemini or Ollama model receives the pixels alongside your text. PNG, JPEG and WebP are supported; convert SVG/BMP or scanned PDF pages to one of those formats. GIF is also sent to providers other than Gemini, subject to their decoder support.

Default visual-input limits are 4 images per call, 5 MiB per image, 10 MiB total; newer messages take precedence. Files can still be stored and downloaded when they exceed these limits. The model receives a notice when an image cannot be supplied. Agents can reopen task-local images with filesystem_read. See Model settings to enable a custom vision model.

Keeping a conversation manageable

  • Pin chat puts it in your personal Pinned list on the Tasks page (up to 50 chats).
  • Pin under an answer saves it above the conversation. Select text for Pin selection; pinned items can be copied, quoted or jumped to (up to 100 per chat, 4,000 characters per selection). These bookmarks do not make model context permanent.
  • Compact asks the root agent to summarize older history. Your visible chat and files stay; a marker records completion. Summaries use the fast model and count toward usage; large histories can need several calls. A request on a busy agent waits for its current turn to end.
  • Pause / Resume appears beside Stop while the team works. Reading earlier messages no longer automatically scrolls you to the bottom; Jump to latest takes you back.
  • Helpful / Not helpful sends a rating to the server's operators. Use a follow-up to change the answer itself.

A follow-up during a run tops up spending allowances to leave a full round, without stacking unused rounds. It does not restart the running task's clock. Reopening a finished task starts a new round. Server defaults are 25 million tokens, $20, 1,000 tool calls and 8 children per round; your task's selected limits, plan and runtime limits still apply.

From code: POST /api/v1/tasks with a goal, then GET /api/v1/tasks/{id}/wait and …/result. You can also start tasks over MCP, A2A or ACP (see Integrations).