Create your first agent

Create an agent when a named, persistent role should own work across multiple check-ins. The agent receives a home project and one continuing conversation, so configure it as an ongoing teammate rather than a scheduled copy of a prompt.

1. Create the profile

Open Agents, select New agent, and enter:

• Name — recognizable in the roster and conversation;
• Persona — optional professional role or operating posture; and
• Description — optional display copy for the agent directory. It is not sent to the model.

Select Create. d5s creates the agent, its managed home project, and its lifetime session, then opens the agent page.

2. Write standing instructions

Open the agent, select the Manage agent settings icon in the breadcrumb, and choose Instructions. Edit Persona and Standing instructions there. Standing instructions should describe the responsibility that remains true across every turn:

• the outcome the agent owns;
• sources it should check;
• decisions it may make;
• actions that require approval;
• the expected output or destination; and
• what to do when there is nothing meaningful to report.

Avoid placing today's task or a one-off deadline in standing instructions. Send that as a direct message, or set up a task-specific scheduled wakeup for it — either by asking the agent or by adding one yourself under Manage agent → Autonomy.

3. Add durable context and capabilities

Use the Manage sections to configure the agent's home project:

• Files for reference material that should remain available;
• Connectors for workspace credentials the agent may use; and
• Skills for reusable methods or instructions.

Activation makes a connector available to the agent. MCP tool policy still controls whether individual operations allow, ask, or deny; API operation controls apply only when they are available and enabled for that credential.

4. Test in the conversation

Keep autonomy off and send a representative request directly to the agent. Review the sources it selected, tool calls it made, approval requests it raised, and whether the result followed the standing instructions.

Test at least three conditions: a normal case, incomplete evidence, and a case where there is nothing to do. Correct the instructions or capability set and repeat until all three behave predictably.

5. Add autonomy last

Open Manage agent → Autonomy only after direct tests pass. You can configure heartbeat frequency, time, timezone, daily budget, and pause state, and you can create, edit, and remove task-specific scheduled wakeups there alongside the ones the agent created.

Start with a conservative cadence and a finite daily budget. A heartbeat asks the agent to inspect its project and decide whether work is due; it still invokes the configured model. Pausing the agent blocks work without deleting its conversation or configuration.

Read Agent wakeups and heartbeats before enabling the first autonomous check-in.