Build an agent from scratch

Create a Worker agent

From the Agents page, click + Agent, name it (e.g. "Invoice Processor"), pick Worker Agent, and create it. A Worker agent is autonomous and queue-driven: it processes one work item at a time with no one watching.

Write the runbook

The runbook is the agent's instructions for handling a single work item, unattended — write it like an SOP. For the Invoice Processor:

• Objective — code each incoming invoice to the right GL accounts and cost centers.
• Steps — extract the invoice, match the vendor, look up how similar invoices have been coded, and assign a GL account and cost center to each line.
• Outcome rules — the most important part of a Worker runbook (see below).

Manage the work-item status

Every work item moves through a tracked lifecycle — Pending → Executing → Completed / Error / Needs review — and your runbook decides where each one lands:

• Complete it when the work is done and the agent is confident.
• Send it to Needs review when a human should decide — a new vendor, an anomalous amount, a possible duplicate, or a low-confidence result. The item pauses; a person can fix it and continue it, or restart it.
• Error is for something the agent genuinely can't handle.

Spell these rules out explicitly — e.g. "if the vendor isn't in the history, flag for review; never auto-code a new vendor." A good Worker runbook is mostly about when to act vs. when to escalate.

Connect data

Attach a semantic data model so the agent can reason over your data in natural language. We'll reuse the shared Meridian AP data from the demo — the model over vendors, invoices, invoice_lines, gl_accounts, and cost_centers. From the agent, open Configure Data Models, connect to the AP data, and attach the model. Now the agent can match vendors and pull historical coding without anyone writing SQL.

The agent's runbook, data, tools, and schedule all live in its config panel:

Add tools (MCP servers)

Attach MCP servers for any actions the agent needs — for example posting a summary to Slack or writing results back to a system.

Test with work items

Create a few work items to run it end to end:

• Upload a file — each file becomes a work item, or
• Create work items via the API.

Watch them move through the lifecycle on the agent's dashboard — the ones that completed, and any that landed in Needs review for you to resolve. Open a work item to see exactly what the agent did. Iterate on the runbook until the outcomes are right.

Create a schedule

For hands-off operation, add a schedule so work items are created automatically on a cadence — for example every morning. Each scheduled run creates a work item the agent picks up, turning the agent into an unattended, recurring process. Give the scheduled work item a name, an optional message and payload, and set how often it runs.