Skip to content

Ask an agent to schedule something ​

You can ask an agent, in any chat you already use with it (Telegram, WhatsApp and so on), to do something on a regular basis. For example: "every Monday at 10, check the open invoices and tell me which are late".

The agent turns that into a schedule, a job AgentX runs at set times. Nothing it asks for runs until you approve it.

Before you start ​

  • The agent can use the AgentX tools. Codex and OpenCode agents always can. A Claude Code agent can when its mcp settings include the AgentX server (agentx serve --stdio).
  • Approval requests have somewhere to go: set notifications.destination in agentx.json (see Get notified). Without it, requests still wait for you, but nobody tells you about them.

1. Ask for it in chat ​

  1. Chat app: tell the agent what to do and when, in plain words. For example: "every weekday at 6pm, summarise today's support tickets and send them here".
  2. The agent replies that it has asked for the schedule and that it's waiting for your approval.

Timing can be written the way you'd say it: "every morning at 9", "weekdays at 6pm", "every 15 minutes", "1st of every month at noon". To see how AgentX reads a phrase:

sh
agentx schedule parse "weekdays at 6pm"

By default, the job runs as the agent you asked, and after each run its answer is posted back to the chat you asked from. If a run fails, a short failure message goes to the same chat. You can ask for another agent, a time zone, somewhere else to send results, or no chat at all ("don't send me the results"). You can also ask to hear only about problems ("tell me only if something is wrong"): then a run that finds nothing wrong sends nothing. See Report only when something is wrong. Without a chat, the results stay in the run history in the dashboard.

If you asked from the AgentX phone app, the answer appears in the same conversation, marked unread. If you asked by voice, it arrives as a push notification on your phone (this needs push notifications set up on this computer). A long answer is shortened to fit the notification; ask the agent by voice and it reads you the full answer.

If you asked from a GitHub or GitLab issue, each answer is posted there as a comment. On a public project anyone can read it. Ask for a private chat instead, or for no chat at all.

If your old schedule prompt already says "send it to my Telegram", remove that line. Otherwise the answer arrives twice: once from the agent, once from the schedule.

In agentx.json, this shows up on the job as its notify destination. See Configuration: automation.

2. Approve or reject it ​

You get a message with the job's timing, the next time it would run, the agent, where the results go, and what it will do.

  1. Terminal: on the machine running AgentX, list the schedules. Waiting requests are marked ◐.
    sh
    agentx schedule list
  2. Approve it, using the ID from the list:
    sh
    agentx schedule approve <id>
    Or turn it down:
    sh
    agentx schedule reject <id>

The change takes effect straight away; no restart needed.

You can also answer from Needs you on the dashboard's Hub, or with agentx approvals list, where requests from agents, memory and the wiki wait together. See Approvals.

A request that isn't approved never runs, even if something else switches it on. Only approve releases it.

What the agent can do ​

Ask forWhat happens
A listIt shows your schedules, whether each is on, who made it, and when it runs next
A new scheduleSaved switched off, marked as waiting, and you're told
Pause or resumeTakes effect at once, for schedules the agent itself made and you approved
RemovalWaits for your approval; the schedule keeps running until then

An agent can only pause, resume or ask to remove schedules it made. To let an agent manage every schedule, set "admin": true on it in agentx.json. Even then, only you can approve.

If an agent withdraws its own request before you approve it, the request is simply removed, since it never ran.

Check it worked ​

  1. Terminal: run agentx schedule list. Your new job is listed without the ◐ mark and with a next run time.
  2. Browser: open the dashboard. On the Hub, select Schedules in the nav, under Automation. The job has a row under its machine, marked on, with its schedule and next run. See See recurring jobs.
  3. After its first run, the agent's answer arrives in the chat you asked from.

If something is wrong ​

  • The agent says it can't schedule things: it doesn't have the AgentX tools. Add the AgentX server to its mcp settings, then restart the daemon (agentx daemon restart).
  • You never got the approval message: set notifications.destination in agentx.json. The request is still waiting in agentx schedule list.
  • approve says the job isn't waiting: it was already approved or rejected. Check agentx schedule list.
  • The agent can't pause a job: it didn't make that job. Make the change yourself, or give the agent "admin": true.
  • The job ran but no answer arrived in the chat:
    1. Terminal: run agentx schedule list. The job's notify: line should name the chat and say (results and failures).
    2. If it says (failures only), open agentx.json and remove "deliverResult": false from the job.
    3. If there is no notify: line, the request was made from a place that isn't a chat (for example another schedule). Add a notify destination to the job, as in Send a daily report.
    4. A run that ends with an empty answer sends nothing. Check the run on the Hub: select Schedules in the nav, then the job's row, to read the last run's summary.
    5. If the notify: line says voice, check that push notifications reach your phone. Without them a voice result can't be sent, and the daemon log shows result delivery failed.
  • The timing is wrong: check the phrase with agentx schedule parse "<phrase>", then ask the agent to change it.

Released under the MIT License.