Choose a work mode
Decide when a task belongs in a chat, agent, automation, or dashboard.
Start with the behavior the work needs, not the label. All four modes can use project context and agent capabilities; they differ mainly in who initiates the work, how continuity is kept, and what people return to afterward. Workspace connectors are available in their projects by default, and you can turn off an inherited connector for the particular project, chat, or agent that should not use it. Conversationally created agents use only their draft-reviewed connections and do not inherit future workspace credentials; see Create your first agent.
| Need | Choose | Execution shape |
|---|---|---|
| A person steers the next step | Chat | One continuing interactive session |
| A persistent role should notice and act | Agent | One lifetime conversation with manual and autonomous turns |
| A defined job should run at a known time | Automation | A separate session for every run |
| People should return to a maintained artifact | Dashboard | Refresh sessions that publish versioned output |
Use a chat for exploration
Choose a chat when a person should stay in the loop, the goal is still changing, or each next step depends on the last answer. Chats are also the best place to prototype instructions before turning them into repeatable work.
Use an agent for responsibility
Choose an agent when the work has an ongoing owner, a persistent conversation, and standing instructions that should survive beyond one run. An agent can receive direct messages, be woken manually, maintain one-time and recurring wakeups, and participate through connected channels.
If you are not sure which agent to staff first, start with one responsibility you want to hand over. Describe the work in an empty Agents directory turns that responsibility into a named coworker and its first assignment.
Do not choose an agent only to obtain a schedule. If every occurrence should start from a stable job definition with independent history, an automation is simpler.
Use an automation for a defined trigger
Choose an automation when the prompt and trigger are known in advance and each run can be evaluated as a repeatable job. Automations support manual-only, one-time, and recurring schedules. Missed recurring slots coalesce instead of creating a backlog of catch-up runs.
Use a dashboard for a maintained output
Choose a dashboard when the primary product is a refreshable HTML artifact rather than a conversation. The build chat is still available for iteration, but the live version is the object other people return to.
You can begin interactively and promote the stable pattern later. The important boundary is whether the work needs dialogue, durable responsibility, a trigger, or a materialized result.