Tasks
The work surface: describe the outcome, approve the plan, watch it get done.
How a task runs
Plan-first, by design: nothing executes until a person has approved it.
Say what you need and attach anything relevant. The task already carries the project and workspace context.
Multiply proposes a plan as a card. You approve it, adjust it, or reject it before anything runs.
Agent activity shows inline while it works, so progress is visible rather than a spinner.
Output lands beside the conversation. You edit, iterate, and decide what ships.
Judgement
Quick questions get quick answers; bigger jobs are delegated to agents. Either way the plan comes first and a person stays the gate, which is the difference between delegating work and losing control of it.
Short answers first. The longer ones are a walkthrough away.
In tasks, no: the plan comes first and a person approves it before execution. Scheduled workflows you have explicitly set up run autonomously, and their output still lands as a task for review.
Research, drafting, synthesis and production steps. Simple asks come back directly; complex ones are planned and delegated to agents.
Bring a real brief to the walkthrough and see the plan it proposes.