Recipes
A recipe is a worked example, not a new feature. Each one names the actors, the prompts that start the work, the signals to watch in the Work pane, the output to expect, and the mistakes that waste a command.
Recipes
| Goal | Recipe |
|---|---|
| Run two independent research tracks and combine the results | Run parallel research |
| Review one change through separate correctness, security, and test lenses | Run a multi-agent code review |
| Turn an issue into a scoped implementation ticket | Turn an issue into a ticket |
| Give one agent a skill without exposing it to the team | Use an agent-scoped skill |
Before you start
- Complete the quickstart and first team. You need a workspace, a leader, and at least two members.
- Give every agent a main model. A recipe that names an agent without a model cannot produce a result.
- Read Conversations and routing for
$Nameand@path, and Dispatch budgets for the ceilings a recipe can hit. - Open the workspace group chat for multi-agent work. A direct chat stays one-to-one and never routes a mention to another agent.
How to use a recipe
- Read Actors and Before you start and adjust the synthetic names to your own roster.
- Copy an Example prompt into the group composer with
i, then pressEnter. Replace<placeholders>before you send it. - Follow What to watch in Work while the work runs. Open a native session with
owhen a result needs inspection. - Compare the result with Expected result. If it differs, read Pitfalls before you retry.
- Use Next to continue to the neighboring task or the exact reference page.
Expected result
Each recipe produces visible work items, a durable answer in the group chat, and a finished history you can inspect. You can name the actor behind each result and the Work state it passed through.
Next
- Run parallel research: the first multi-agent recipe.
- Guides: the single-agent task guides behind these recipes.
- Concepts: the mental model for actors and Work.