Automations

Rule

Automations are the user-facing umbrella for agent-executed tasks that fire in response to events or on a cron schedule. Scheduled and Event are the two trigger types. Each automation is a markdown resource under jobs/ with YAML frontmatter describing when and how it fires, and a body containing natural-language instructions the agent follows.

Recurring Jobs is the legacy name and API for scheduled automations. manage-jobs, jobs/, and /agent#jobs remain stable compatibility surfaces. Use manage-automations for new personal or organization automations, including per-automation model overrides and MCP allowlists. Keep manage-jobs for existing schedule-only integrations and delivery metadata.

The Two Trigger Types

Event triggers can optionally include a condition -- a natural-language string evaluated by Haiku against the event payload before dispatch. If the condition does not match, the automation is skipped.

How It Works

1. User asks the agent to create an automation (or uses the app's Automations page in Settings, /settings/automations; open it with open-settings-page page automations).
2. Agent calls manage-automations with action=list-events to discover available events.
3. Agent calls manage-automations with action=define to write a jobs/<name>.md resource.
4. The trigger dispatcher subscribes to the event on the bus.
5. When the event fires, the dispatcher loads all matching triggers, enforces owner and organization scope, and evaluates conditions via Haiku.
6. Event and cron acquisition converge on the shared background-automation runner, which validates identity, resolves the configured model and MCP allowlist, runs the agent loop, handles continuation and delivery, and records usage.
7. Status (lastRun, lastStatus, lastError) is written back to the resource frontmatter.

Trigger acquisition stays separate by design: the scheduler decides when a cron expression is due, while the event dispatcher matches event names, owners, and conditions. Everything after a trigger is accepted uses the same execution lifecycle.

Markdown Format

---
schedule: ""
enabled: true
triggerType: event
event: calendar.booking.created
condition: "attendee email ends with @example.com"
mode: agentic
domain: calendar
createdBy: user@example.com
runAs: creator
---

Send a Slack message to #sales with the booking details.
Use the web-request tool with ${keys.SLACK_WEBHOOK}.

Frontmatter Fields

Agent Tools

All automation operations are accessed through a single manage-automations tool with an action parameter:

Additional tool: web-request — outbound HTTP with ${keys.NAME} substitution.

manage-automations accepts personal or organization scope and supports model, reasoning_effort, and mcpTools on define/update. An MCP allowlist is enforced, not advisory: every named tool must resolve in the creator's request context or the run fails clearly, and the runner never widens access beyond the configured names.

Organization Event Automations

Organization event automations are visible to organization members but always run as their creator:

• createdBy is required and runAs must be creator.
• An organization event automation matches only events emitted for that creator. Organization visibility does not turn an event into a broadcast.
• Organization admins may update or delete an automation, but cannot replace its creator or retarget its execution identity.
• Before every run, the framework verifies that the creator still exists and is still a member of the organization. A removed creator or unreadable membership state prevents execution.

The Event Bus

Integrations register events at module load time. The bus validates payloads against Standard Schema definitions and dispatches to subscribers.

import { registerEvent, emit } from "@agent-native/core/event-bus";
import { z } from "zod";

// Register an event type (typically in a server plugin)
registerEvent({
  name: "calendar.booking.created",
  description: "A new calendar booking was created",
  payloadSchema: z.object({
    bookingId: z.string(),
    attendeeEmail: z.string(),
    startTime: z.string(),
  }),
  example: { bookingId: "abc", attendeeEmail: "jane@co.com", startTime: "2025-01-15T10:00:00Z" },
});

// Emit the event (from an action, webhook handler, etc.)
emit("calendar.booking.created", {
  bookingId: "abc",
  attendeeEmail: "jane@co.com",
  startTime: "2025-01-15T10:00:00Z",
}, { owner: "user@example.com" });

Built-in Events

Event Bus API

Condition Evaluator

When an automation has a condition, the dispatcher calls the configured fast/classification model to classify whether the event payload satisfies the condition. This is a yes/no classification, not a generation task. The exact model ID lives in condition-evaluator.ts.

• Empty or missing condition = unconditional (always fires).
• Results are memoized (SHA-256 of condition + payload) with a 5-minute TTL and 500-entry LRU cache.
• Payload is truncated to 4000 characters before sending to Haiku.
• On API failure, the condition evaluates to false (safe default -- skips the automation).

The web-request Tool and Keys

Automations use the web-request tool for outbound HTTP. It supports ${keys.NAME} placeholders in the URL, headers, and body. These are resolved server-side after the agent emits the tool call -- the raw secret value never enters the agent's context.

• Keys are ad-hoc secrets created by the user on Settings › API keys or through the /_agent-native/secrets/adhoc API.
• Each key can have a URL allowlist that restricts which origins the key can be sent to.
• resolveKeyReferences() resolves placeholders, falling back from user scope to workspace scope.
• validateUrlAllowlist() checks the resolved URL against per-key allowlists (origin-level matching).
• Automation definitions, examples, event payloads, and prompts must not hardcode real API keys, webhook URLs, tokens, private Builder/internal data, or customer data. Use ${keys.NAME} and synthetic example.com identities.

Failure Handling

An automation fails once, with its real cause, instead of re-failing every tick. The runner classifies each failure (jobs/automation-outcome.ts):

• An event or webhook automation counts a failure once per event (lastFailedEventId): queue retries of the same event never pause it alone.
• Preconditions are checked before a thread or agent_runs row exists: no "Job:" thread per tick. The failure is recorded on the automation and its run history only. Never write a generic "ended with status: errored"; surface the run's own error.
• A paused automation has enabled: false, lastStatus: paused, and pausedReason/pausedAt. The owner is emailed once by the run that paused it (for organization and shared jobs: the creator while still a member, otherwise an org owner or admin). Enabling the automation clears the pause and the streak. One paused for an absent credential (missing_credentials) resumes by itself once its identity has a usable LLM credential; a rejected key stays paused until the owner enables it again.
• Transient failures (spent credits, an unreadable credential store, a remote host, 429/5xx) pause like runtime ones, but the scheduler lifts the pause for one probe run on the same backoff (at most 6h). A failing probe pauses again at once without a new email; a successful one clears the streak.
• A paused job's "Run now" is still settled (stale running, remote reconcile); it never moves the streak, and resumes the job only on success.
• Pauses and automatic resumes are measurable: automation_paused (error_code, failure_kind, consecutive_failures, surface) and automation_resumed (via) carry automation_hash, the first 12 hex characters of sha256 of <app>|<name>, never the name. A failed run's capture carries its own thread, run and automation name in extra.failureContext.
• runAs: shared and organization jobs run as the organization, never as the creator's personal connection: only an org admin can connect the provider.
• Jobs owned by users that no longer exist, or by reserved test identities (.test, .invalid, .example, example.com|org|net) in production, are disabled with owner_missing / owner_reserved. Absence is proven only by a populated built-in "user" table (case-insensitive); a lookup error, or a deployment whose accounts live elsewhere (custom getSession), is never read as "deleted".
• The scheduler tick also closes automation_runs stuck running for over 25 hours (interrupted, automation_run_abandoned) and A2A tasks idle for 24 hours (failed, a2a_task_abandoned), a bounded batch at a time.

UI

The full-page Agent surface's Automations tab is the primary management surface for scheduled and event-triggered automations. Users can view status, enable/disable, inspect, and delete automations there. Its URL remains /agent#jobs for compatibility even though the visible tab is Automations. Creation typically happens through the agent chat.

Example

User: "When someone books a meeting with a @example.com email, message me in Slack."

Agent flow:

1. Calls manage-automations with action=list-events to find calendar.booking.created.
2. Confirms the plan with the user.
3. Calls manage-automations with action=define:
   • name: slack-on-example-booking
   • trigger_type: event
   • event: calendar.booking.created
   • condition: attendee email ends with @example.com
   • mode: agentic
   • domain: calendar
   • body: Send a Slack message to #sales with the booking details. Use the web-request tool to POST to ${keys.SLACK_WEBHOOK}.

Key Files

Related Skills

• recurring-jobs -- schedule-triggered automations reuse the same scheduler
• secrets -- ad-hoc keys and ${keys.NAME} substitution
• actions -- automations can call any registered action via the agent loop
• delegate-to-agent -- agentic mode runs a full runAgentLoop