You can build every automation in autobotAI node by node, but most teams start by describing the job:
- AI Builder works on the workflow canvas. Describe a workflow, or a change to one, and it builds it as nodes you can read, test and edit.
- Optimus works at the level of goals. Describe an outcome, and it assembles a harness: the agents, triggers, memory and approvals the goal needs.
Either way, the result is ordinary building blocks (bots, listeners, schedules, memory spaces) that you can open, change, version and export.
Build a workflow with AI Builder
- Go to Action Resources > Bots and create a bot. The canvas opens with the AI Builder panel.
- Describe the job end to end: what triggers it, what it checks, what it changes and who gets told. For example: "Every day, find IAM access keys older than 90 days, look up each key's owner, ask the owner's manager in Slack to approve deactivation, then deactivate approved keys and post a summary."
- Answer once. AI Builder asks once for anything it needs, such as recipients, approvers or a schedule.
- Review the canvas. Each step appears as a node you can open. Check that there is an approval node before any step that changes something.
- Change one part at a time. Select a node and describe the change, such as "add a condition after this step that skips service accounts". AI Builder updates the flow at that step.
- Ask for an explanation at any point: "Explain what this bot does, step by step."
- Test, then enable. Run it manually, check the execution, then switch it on.
AI Builder works on the saved bot. When a bot is already enabled, you are asked to save or discard canvas edits before AI Builder runs, so nothing live changes unexpectedly.
Write a good brief
| Include | Example |
|---|---|
| The trigger | "every Monday at 06:00", "when GuardDuty raises a finding", "when someone requests access in the service portal" |
| The scope | "production accounts only", "the payments-api repository" |
| The decision | "if the key was used in the last 7 days, skip it" |
| Who approves | "the resource owner's manager, in Slack" |
| What to change | "deactivate the key", "open a pull request", "add a WAF rule in count mode" |
| Who is told, and how | "post a summary to #secops; email the owner" |
Build a harness with Optimus
Use a harness when the outcome spans many runs, tools and owners. Open Optimus and describe the goal and its boundaries, for example:
"Keep checkout-service free of exploitable vulnerabilities. Scan it nightly, take findings from Snyk and SonarQube, deduplicate them, work high-severity findings toward a fix as a pull request approved by the owning team, re-test fixed findings every six hours and keep the risk register current."
Optimus plans the harness and creates its parts:
| Part | What Optimus sets up |
|---|---|
| Triggers | Schedules and listeners, such as a nightly scan and a webhook for findings from other tools |
| Agents | Workflows and deep agents, each with its job, tools and model role |
| Memory | Memory spaces for findings and shared context, so agents recall earlier runs |
| Approvals | Approval steps with named approvers in Slack, Microsoft Teams or the pull request |
| Proof | A re-test agent and updates to the risk register |
Review each part before you switch the harness on. Every agent it creates is a normal bot you can open and edit, and every trigger is a normal listener or schedule.
Keep control
- Tools are opt-in. AI Builder and Optimus only use integrations and actions you have connected and enabled.
- Approvals are explicit. Ask for an approval node before every change, and check that it is there.
- Autonomy is yours to raise. New automations should start at advisory or gated actions; raise an action's autonomy only once its record shows it is safe. See Autonomy and approvals.