Example workflows

These examples assume the agent is connected over MCP or REST. Tool names are model__command over MCP and /api/model/command over REST.

Triage new helpdesk tickets ​

Needs: the Helpdesk package installed. Scopes: helpdesk.agent.

Prompt

You help the support team. List helpdesk tickets in state open that have no category. For each, read the subject and description, set category (question, incident, problem or feature_request) and priority (low, normal, high or urgent) and give me a one-sentence summary in your final report. Do not change the state or any other field. Personal data is masked; never try to find out who the customer is. At the end, list what you changed.

What the agent does

1. helpdesk_ticket__search with {"filter":[{"field":"state","op":"eq","value":"open"},{"field":"category","op":"is_null"}],"limit":25}.
2. For each row, helpdesk_ticket__update with {"id":"...","category":"incident","priority":"high"}.
3. Reports the list.

Where a person steps in: replying to customers. Helpdesk has no sending action, and this agent has no way to contact anyone.

Draft a quotation ​

Scopes: crm.user, sales.user, base.user, product.user.

Prompt

The lead "Warehouse pilot" has reached the qualified stage. The customer is the party with id 01a1464f-b0c8-700e-9f39-0cbf8e848c0b. Look up the list prices of the products I name below and create a draft sale order for that party with the matching total and currency EUR. Leave it in draft. Tell me the order id.

What the agent does

1. lead__search to confirm the lead and its stage.
2. product__search for list prices (list_price_minor, an integer in minor units).
3. sale_order__create with party_id, total_minor and currency. The state starts at draft and the agent cannot set it to anything else.

The prompt gives the party id because masked contact details on a lead cannot be matched to a party: tokens differ per field, and an agent cannot search masked fields.

Where a person steps in: confirming the order. The agent could run sale_order__confirm too, because Sales has no approval step; if you want a person to confirm, do not put confirm in the prompt, or keep the agent's scope to a read-only group.

Prepare invoices for approval ​

Scopes: invoicing.user, invoicing.manager.

Prompt

List invoices in draft. For each, check that total_minor is greater than zero and that it refers to a sale order. Propose posting the ones that pass and give me a table of invoice id, total and plan id. Do not try to post anything yourself.

What the agent does: invoice__search for drafts, then invoice__post for each that passes. Every call returns approval_required with a plan id. The agent lists the plans. A person with invoicing.manager approves the ones they agree with, using the approval loop. The legal number is assigned only when the plan runs.

Keep a CRM pipeline tidy ​

Scopes: crm.user.

Prompt

Find leads in qualified with no expected revenue and list them for me. Do not edit them. Then find leads in new that have no contact name and suggest which to archive. Do not delete anything; if you think a lead should go, say which and why.

An agent can only propose a delete, so even a mistaken instruction does no immediate harm. The plan appears as approval_required.

A scheduled daily summary over the CLI ​

Scopes: read-only groups such as crm.user.

export OBOP_TOKEN=obop_read_only_token
obop call lead search 'filter=[{"field":"stage","op":"eq","value":"won"}]' limit=50 > won.json

A script feeds won.json to a model to write the summary. Because leads come back masked, the summary refers to tokens, not names. Resolve names yourself afterwards if you need them.

Writing good agent instructions ​

• State the goal and the boundary: what to do, and what never to do.
• Name the states and fields you mean (state, stage, category), so the agent does not guess.
• Say what to do when something is withheld or needs approval: report it and stop.
• Ask for a short report at the end. It is what the approving person reads.

See Safety best practices and Troubleshooting.