Agent-Managed Automations

Ask the agent, in chat, to set up and manage your automations for you — draft a config from a description, create a scheduled or event-triggered automation, edit or pause one, or delete it. Every change runs with your identity and is gated behind the same human approval card you already use for sensitive actions, so a person always signs off before an automation is created, edited, or removed.

This is a conversational front door to the automation wizard — it does not bypass it. The agent proposes; you approve.

What you can do

Capability	Tool	Approval
Draft a config from a description	suggest_automation_config	None — it's a read-only draft, nothing is saved
Parse a schedule ("every weekday at 9am") into a cron expression	parse_cron_expression	None — read-only
List / inspect your team's automations	list_automations, get_automation	None — read-only
Create a scheduled / event-triggered automation	create_automation	Required
Update an automation (edit fields, pause/resume via is_enabled)	update_automation	Required
Delete an automation	delete_automation	Required

The two draft tools (suggest_automation_config, parse_cron_expression) let the agent turn a plain-language request into a concrete proposal — a suggested system prompt, the instruction the automation runs each time, a shortlist of relevant tools, and a valid cron schedule — before it proposes the actual create_automation. If your request is too vague to draft confidently, the agent asks you for the missing detail (what triggers it, what it should do, what "done" looks like) instead of guessing.

Turning it on

Agent-managed automations are off by default. A team admin enables them per team (the agent_automation setting, an off-by-default per-team opt-in that mirrors the code-execution sandbox). Until the team is opted in, none of the tools above appear — the agent simply won't offer to manage automations, and a forced request fails closed without writing anything.

When you enable it, the whole family turns on together: the draft helpers, the read tools, and the approval-gated write tools.

Where it works

The write tools are available only in research mode; the read and draft tools ride the default chat tier too.

Chat mode	Draft + read tools	Write tools (create / update / delete)
Default / auto	Available (opted-in team)	Not exposed
Research	Available	Available, each behind an approval card

So if you want the agent to actually create or change an automation, switch the conversation to research mode. In default chat the agent can still draft a config, list your automations, and inspect one — it just can't write.

Permissions and governance

• You are the owner. An automation the agent creates on your behalf is stamped created_by = you. Its permissions and RBAC are your own — the agent is acting as you, not as a separate service identity.
• Every write needs approval. create_automation and update_automation are risk tier T4; delete_automation is T5. All three pause the run and surface an approval card showing a human-readable summary of what will change (for an update, a before → after diff; for a delete, the name and how many scheduled / in-flight runs will be cancelled). The automation is created, edited, or removed only after you approve.
• skip_approvals is not settable via the agent. Bypassing approvals is an owner/admin-only setting changed in the automations UI — the agent tools refuse it outright, so the approval gate can't be laundered through chat.
• Teams with custom risk thresholds: a team can tighten governance with a risk-tier deny threshold, which hard-denies every tool at or above that tier (no approval path). With the usual threshold this denies only delete_automation (T5) while create_automation and update_automation (T4) still reach you through the approval card; a stricter threshold set at T4 hard-denies create and update as well. Manage any hard-denied automation from the automations UI instead.

See Human Approval for how the approval card is routed and decided.

Cost

Creating, editing, or deleting an automation is free — drafting and the write itself are lightweight operations. What's charged is the automation running: every scheduled or triggered run is billed against that automation's budget, with the same per-run and per-team-day cost ceilings as any agent run. See Cost & Auto-Resume for how those budgets and caps work.

Deleting an automation

Deleting an automation is a soft delete: it's hidden from your active automations and stops firing, but its run and cost history is preserved for audit. The automation row is never physically removed, so nothing cascades away — you just won't see it in the automations list anymore, and there's no self-serve undo from the UI.

delete_automation is risk tier T5 and, like every write, pauses the run for your approval — the card shows the automation name and how many scheduled / in-flight runs will be cancelled. If you might want the automation back, prefer update_automation with is_enabled=false to pause it instead: a paused automation stays in your list and resumes with one approval.

Walkthrough

1. Switch the conversation to research mode.
2. Ask for what you want, in plain language — for example: "Every weekday at 8am, summarize new support tickets and post the summary to our team."
3. The agent drafts the config (it may call suggest_automation_config and parse_cron_expression behind the scenes) and proposes create_automation. The run pauses and you see an approval card with the summary: name, trigger (cron 0 8 * * 1-5), the instruction it will run, and the tools it will use.
4. Review the summary and click Approve. The automation is created (owned by you) and, for a cron trigger, scheduled for its next run.
5. From then on it runs on its schedule. To pause it later, just ask: "pause the daily ticket digest" → the agent proposes an update_automation with is_enabled=false, which you approve the same way.

Troubleshooting

Symptom	Likely cause
The agent says it can't manage automations, or the tools never appear	The team isn't opted in. A team admin must enable agent-managed automations for the team.
The agent answered but never showed an approval card for a create / update / delete	You're in default / auto chat, which only exposes the read and draft tools. Switch to research mode to make changes.
You approved a create but the automation never fires	Check it's enabled and the schedule is valid — an invalid cron expression won't schedule. Also check whether a run hit a cost or budget cap and is paused.
A create / update / delete is refused with no approval option	Your team has a risk-tier deny threshold that hard-denies tools at or above that tier (delete is T5; create and update are T4). Manage those automations from the automations UI.

Best practices

1. Use research mode when you intend to change anything; stay in default chat for "what automations do I have?" or "draft me a config" questions.
2. Read the approval card before approving — it's the summary of exactly what will be created or changed, including the trigger and the tools the automation may use.
3. Pause (is_enabled=false) rather than delete when you're unsure — a paused automation stays in your list and resumes with one approval, while a deleted one is hidden from view (its run history is retained for audit, but it isn't restored from the UI).
4. Be specific in your request. "Make me an automation" forces the agent to ask follow-ups; "every Monday at 9am, email me last week's open tickets" lets it draft a complete proposal in one turn.

Related

• AI Agents — what an automation is (triggers, delivery, acceptance criteria)
• Human Approval — how the approval card is routed and decided
• Cost & Auto-Resume — the budgets a running automation is charged against