Use auxilia from n8n & Cloud Scheduler

Agents are most useful when they don’t wait for someone to open a chat window. This tutorial shows two common ways to put them into motion automatically: automation platforms (n8n, Zapier, Make) and scheduled jobs (Google Cloud Scheduler, cron).

Both use the same underlying mechanism — a simple HTTP request — so if you haven’t already, skim Trigger agents over HTTP first to understand the shape of the call.

Pattern 1 — Automation workflows (n8n, Zapier, Make)

Automation tools are ideal for event-driven agents: when something happens in system A, ask an agent to do something about it, then push the result into system B.

Example: auto-respond to HubSpot form submissions

Goal: when someone fills in a contact form, have an agent draft a tailored response and post a suggestion in Slack for a sales rep to review.

The workflow nodes:

1. Trigger — HubSpot “New form submission”
2. Action: HTTP Request — POST to https://your-auxilia-host/api/backend/threads/runs/invoke
   • Header: Authorization: Bearer <your PAT>
   • Body:

     {
       "agent_id": "your-crm-agent-id",
       "message": "A prospect filled out our contact form. Name: {{contact.firstname}} {{contact.lastname}}, company: {{contact.company}}, question: {{form.message}}. Draft a short, friendly response and flag the account priority."
     }

3. Action: Slack “Post message” — post the agent’s response to #sales-inbound

The whole thing runs in a few seconds and scales however your automation tool scales.

Good candidates for this pattern

• New support ticket → agent summarizes and suggests a reply
• New Stripe customer → agent enriches the record from their website and writes it to Salesforce
• Failed CI run → agent reads the logs, summarizes likely cause, comments on the PR
• Calendar event ends → agent pulls notes from the doc and updates the CRM

Tips

• Give the agent structured data in the prompt — JSON-like text, template variables. The LLM handles it fine.
• Use tool approvals carefully — if this workflow runs unattended, the agent’s tools need to be on always allow, or the run will pause waiting for a human
• Start with a non-destructive agent — one that drafts and suggests, then graduate to one that writes back to systems once you trust the output

Pattern 2 — Scheduled jobs (Cloud Scheduler, cron)

Schedulers are great for recurring agent runs: daily summaries, weekly reports, monthly audits. The agent runs on its own clock and posts results somewhere useful.

Example: daily pipeline report in Slack

Goal: every weekday at 8am, have the CRM agent summarize yesterday’s deal activity and post it to #sales-stand-up.

Using Google Cloud Scheduler:

1. Create a Cloud Scheduler job with schedule 0 8 * * 1-5 and timezone Europe/Paris
2. Target: HTTP
   • URL: https://your-auxilia-host/api/backend/threads/runs/invoke
   • HTTP method: POST
   • Auth: Add header Authorization: Bearer <your PAT>
   • Body:

     {
       "agent_id": "your-crm-agent-id",
       "message": "Summarize yesterday's deal activity from HubSpot — new deals, stage changes, closed-won. Post the summary to #sales-stand-up with a one-line highlight at the top."
     }

If the agent has the Slack MCP server bound and the right tools on always allow, the posting happens inside the agent — you don’t need a second workflow step.

Good candidates for scheduling

• Daily CRM / pipeline digest
• Weekly “what shipped” summary from GitHub + Linear
• Monthly cost roll-up from billing MCP servers
• Quarterly KPI report to a leadership Slack channel

Tips

• Think about time zones — schedulers normally run in UTC; set the zone explicitly or the 8am job fires at a surprising hour
• Use idempotent agents — if the scheduler retries on failure, you don’t want duplicate posts. Tell the agent to check for an existing post before creating a new one.
• Monitor through Langfuse — scheduled runs don’t have a human watching, so observability is how you catch regressions

Combining both

Nothing stops you from mixing patterns. A common one:

1. Scheduler wakes up every morning and asks an agent to check for overdue invoices
2. Agent finds the list and calls a webhook (via a custom MCP tool or HTTP MCP server) into your n8n workflow
3. Workflow does the fan-out: one Slack message per overdue invoice in the right channel, one email per finance contact

This way the agent’s “intelligence” lives in auxilia (prompt, tools, model choice) and the “plumbing” lives in n8n — each tool doing what it’s best at.