---
title: "Run your first process"
description: "This guide takes one real business procedure from a written description to a monitored run in sandbox. You will choose a low-risk process, describe its steps and outcome, and assign an agent with a narrow tool scope. You then run it against recorded inputs and compare the result with the outcome you declared."
url: "https://aeterion.tech/en/docs/guides/run-your-first-process"
locale: "en"
publisher: "AETERION SA"
---
# Run your first process

> This guide takes one real business procedure from a written description to a monitored run in sandbox. You will choose a low-risk process, describe its steps and outcome, and assign an agent with a narrow tool scope. You then run it against recorded inputs and compare the result with the outcome you declared.

Source: https://aeterion.tech/en/docs/guides/run-your-first-process
Last updated: 2026-09-20

## Choose the right process

The first process should be frequent enough to produce signal within days and contained enough that an error is recoverable.

- Runs at least fifty times a month.
- Has a written outcome someone can verify in under a minute.
- Touches two systems at most.
- Has an owner who can approve escalations during working hours.

## Describe it as a definition

Write the steps as they happen today, including the checks a person performs. Those checks become policies later; do not drop them at this stage.

`vendor-onboarding.yaml`
```yaml
name: vendor-onboarding
owner: procurement
trigger:
  type: form
  source: procurement.vendor_request
steps:
  - id: verify-registry
    agent: research-agent
    tools: [registry.company_lookup]
  - id: check-sanctions
    agent: risk-agent
    tools: [compliance.sanctions_screen]
  - id: create-record
    agent: procurement-agent
    tools: [erp.vendor_master.write]
outcome:
  success: vendor_record_created
  sla_minutes: 120
```

## Run against recorded inputs

Replay real requests from the past weeks instead of inventing test cases. Recorded inputs expose the exceptions that written procedures rarely mention.

```bash
braian process run vendor-onboarding --replay ./requests-august.jsonl
```

## Verify before you promote

- Every completed run reaches the declared outcome, or escalates with a reason you accept.
- No step calls a tool outside the agent's declared scope.
- The audit entries contain enough context for the process owner to reconstruct a decision.

> **Do not promote on volume alone** — A high completion rate in sandbox with synthetic data says little. Promote only after a replay of real inputs reaches the declared outcome.

