Scheduling Work for Agents

An agent can promise to do something later, like "every weekday at 9am, summarize my open pull requests", and keep that promise with a dynamic schedule. eve has this built in. With other harnesses, you write two pieces, a tool and a consumer:

1. A tool creates the schedule. The model calls it with a time and a prompt. The tool calls Schedules.create() with a Edgenivo Queue topic as the target, the user's ID as the namespace, and the prompt in the payload.
2. The schedule fires. Edgenivo publishes a message to the topic with the payload.
3. A consumer wakes the agent. It sends the prompt back to the agent, as if the user had just asked. With a durable harness, it can send the prompt into the same conversation.

• Use a namespace per user. The tool decides the namespace, not the model, so a user can only list and cancel their own schedules. See Multi-tenant schedules.
• Name schedules after the tool call. If a harness retries a tool call, the same name returns schedule_already_exists instead of creating a second schedule.
• Write a standalone prompt. Tell the model to write the scheduled prompt as a complete instruction, and give it the user's current time and time zone so it can resolve "tomorrow at 9".
• Return errors to the model. Return a SchedulesApiError's message from the tool, such as an invalid cron expression, so the model can fix the call.
• Expect retries. Queue messages can be delivered more than once. Use the message's executionId to deduplicate. See Handling scheduled runs.

• Running agent harnesses durably
• SDK reference
• Local development