Build with d5s

Build with d5s after you can describe a useful result and recognize whether it is correct. The safest path is to prove the work interactively, preserve its context, and add autonomy only after the definition is reliable.

Prototype the outcome in a chat

Run the task with representative input. Refine the instructions until the agent consistently identifies the right sources, produces the expected deliverable, and behaves sensibly when information is missing.

Save the chat as a project when its instructions, files, memory, skills, or connector activations should be reused. Keep durable background in the project and occurrence-specific input in the task itself.

Choose the execution shape

Use an agent for an ongoing responsibility with one continuing conversation. Use an automation for a stable job whose runs should remain independent. Use a dashboard when the maintained artifact is the thing people return to.

Do not add a schedule merely because the task repeats. First decide whether the work needs persistent judgment or a repeatable definition. See Choose a work mode.

Add the narrowest capabilities

Activate only the files, skills, and connector credentials required for the job. A workspace connection is not automatically available to every project. Where operation controls are available, keep writes ask-gated until the workflow is trusted. Otherwise use a read-only or narrowly permissioned external account and keep writes interactive.

Test before enabling autonomy

For an agent, send direct messages before enabling its heartbeat. For an automation, keep the trigger Manual and use Run now before adding a schedule. Review the run record, tool activity, usage, and destination—not only the final prose.

Start with Create your first agent or Create your first automation.