Optimus is the autonomous orchestrator in autobotAI. You describe a goal; Optimus plans it, splits it into steps and delegates each step to the right specialist agents, workflows and integrations. Optimus is a swarm, not a single agent: several agents plan, act and check each other's work.
Optimus works inside the same guardrails as every other agent. It can only use the tools you give it, every step is recorded, and actions that need approval still wait for a named person.
Three ways to work with Optimus
Teams usually move through these in order: start ad hoc, automate what repeats, then let trusted work run on its own.
1. Ad hoc, in a chat
Click Optimus in the top bar, describe the job, and the agents do the work now. For example:
- "Here's an alert. Triage it, correlate the logs and suggest containment."
- "Collect the evidence for our quarterly access review and draft the report."
- "Review this IAM policy for privilege escalation paths."
If the work is worth repeating, ask Optimus to save it as an automation from the same session. See Investigate and operate with Optimus.
2. Repeatable workflows
Attach the work to a schedule, an event or a self-service request. Optimus or AI Builder drafts the workflow; you review it on the canvas and tune the mix of deterministic steps and AI judgement. For critical operations, keep the deciding steps deterministic so the result is the same every time.
3. Autonomous, goal-driven work
For outcomes that span many runs, Optimus assembles a whole harness: the agents, triggers, memory, guardrails and checks the goal needs. Deep agents work in maker and checker pairs, and people stay at the checkpoints that matter. Progress shows up in Guardian as tasks, and accepted risks in its risk register.
Not sure whether a job needs a workflow or a harness? See Choosing a workflow or a harness.
What Optimus can use
You configure what Optimus works with:
| Setting | What it controls |
|---|---|
| Model | The model, or models, Optimus uses. See Deployment and models. |
| Tools | The integrations, actions and workflows Optimus may call. Anything not granted is out of reach. |
| Knowledge | Your policies, runbooks and reference material. |
| Memory | Memory spaces Optimus can recall from, so recurring hunts or reviews don't need to be set up again. |
| Skills | Versioned skill packages that teach Optimus how your team does a job. |
| Instructions | A system prompt for house rules, tone and boundaries. |
Skills
A skill packages know-how, such as how your team triages a cloud finding or writes a detection rule, so agents apply it the same way every time.
- Skills are versioned (semantic versions, such as 1.2.0) with notes for each release.
- Each upload is scanned, and a skill that contains a malicious prompt is flagged before it can be published.
- Platform skills can be forked and adapted to your environment.
See Skills for uploading, reviewing scans and publishing.
What teams ask Optimus to do
- Triage an alert, correlate logs and propose containment.
- Write detection rules from a description of attacker behavior, and back-test them.
- Draft a full response playbook from a trigger you describe.
- Schedule a monthly compliance check, scope it and send stakeholders a summary.
- Connect a new security tool, including API setup, credentials and validation.
- Configure alert routing and escalation in Slack, Microsoft Teams or PagerDuty.
- Build self-service security runbooks your wider team can request.
Guardrails that apply to Optimus
- No self-granted autonomy. Optimus can't raise its own autonomy tier or reach tools it hasn't been given. See Autonomy and approvals.
- Approvals still apply. A change that needs approval waits for a named approver, whoever proposed it.
- Everything is recorded. Plans, tool calls, results and decisions appear in the execution history and audit trail.
- You own the output. Workflows Optimus drafts are ordinary workflows you can read, edit, version and export.