Recurring Jobs

Rule

Recurring jobs are scheduled tasks the agent executes automatically on a cron schedule. Jobs live as resource files under jobs/ with YAML frontmatter for scheduling metadata.

How It Works

1. User asks for something recurring via the agent chat
2. Agent uses manage-jobs tool (action: "create") to write a job file at jobs/<name>.md
3. A scheduler polls every 60 seconds, finds due jobs, and executes them via runAgentLoop
4. Job results are saved as chat threads

Connected MCPs in background jobs

Jobs can use connected remote MCPs with the same server-side OAuth lifecycle as interactive chat. When creating a job that needs an MCP, bind the exact advertised mcp__<server>__<tool> names through mcpTools. The scheduler resolves only that allowlist under the job's persisted user/org request context; it never stores or exposes OAuth tokens, URLs, or arbitrary proxy targets. A revoked connector or missing tool fails the run clearly instead of silently widening access.

Use an app-owned bounded import/upsert action for writes. Keep provider-specific response mapping, provenance, deduplication, and write policy in the app rather than in core. For example, a job can read meeting notes from any connected MCP and pass normalized action items to an app's idempotent import action.

Job Tool (built in)

UI Surface

Users can see and manage jobs without the agent on the Agent page's Jobs tab (/agent#jobs, AgentJobsTab in packages/core/src/client/agent-page/): recurring jobs (personal and organization scope) with pause/resume/delete and run status, plus automations (personal-only today). Backed by the scoped list/manage actions in packages/core/src/jobs/actions/ — direct users there instead of describing job files when they just want to view or toggle jobs.

Key Files

Related Skills

• actions — How tools and actions work
• delegate-to-agent — How jobs invoke the agent loop