For years, cloud teams have drawn a clear line when it comes to autonomous optimization: containerized workloads are one thing, but stateful infrastructure is another. Virtual machines and databases carry a different kind of risk — changes are harder to reverse, timing matters enormously, and a misstep can ripple across every application that depends on them.
That's exactly why Sedai is announcing Autopilot for VMs and Amazon RDS — and exactly why we built it the way we did.
From Copilot to Autopilot
Sedai has long supported VMs and RDS in Copilot mode, where our AI surfaces rightsizing recommendations and your team approves each action before it executes. Copilot works well — it significantly reduces the manual toil of researching, sizing, and executing changes — but at a certain scale, the approval process itself becomes the bottleneck. For organizations managing thousands of instances across multiple clouds, asking engineers to click "approve" on every optimization isn't sustainable.
Autopilot closes that gap. Instead of waiting for a human to confirm that now is a safe time to act, Sedai takes on that judgment and executes automatically, within the guardrails you define.
We've supported full Autopilot for Kubernetes for some time. Extending it to VMs and RDS brings autonomous optimization to two of the most widely used (and most carefully managed) resource types in the cloud.
Why VMs and RDS Required a Different Approach
Not all resources carry the same blast radius. A VM resize in a standalone workload is one thing. A database resize that brings down the applications depending on it is quite another.
That's why our team invested significantly in the safety model before releasing Autopilot for these resource types. Two principles guided the design:
- Maintenance windows are required. Sedai will only execute changes within a defined maintenance window. This ensures that any brief disruption during a resize happens at a time your team has already designated as acceptable — not in the middle of peak traffic.
- Optimization is incremental, not a one-shot drop. Rather than jumping from an oversized instance directly to the recommended floor in a single move, Sedai navigates a step-by-step transition path. If you're running an E-series instance and the model recommends a B-series, Sedai moves to D first, stabilizes, re-evaluates, then continues. At each step, if something doesn't look right, the system can pause or roll back — before committing to the next stage.
This is the same reinforcement learning-based approach we use for Kubernetes rightsizing, now applied to VMs and RDS.
How It Works: VMs
Autopilot for VMs is available across cloud providers — not limited to AWS — and covers standalone instances as well as those behind load balancing groups.
When Autopilot is enabled, Sedai's decision engine maps out all possible transition paths from your current instance type to the target state. It then selects the safest path: one that stays within the same instance family where possible, moves one step at a time, and pauses to validate performance at each stage before proceeding.
Sedai determines the target state based on a customer's optimization goals and the guardrails they've established, then handles the journey to get there autonomously, on a timeline calibrated to real-world behavior rather than a single aggressive cutover.
How It Works: RDS
For RDS, Autopilot covers both instance rightsizing and storage optimization — two changes that, in the context of managed databases, are best handled as a coordinated sequence rather than independently.
Sedai first rightsizes the underlying instance (CPU and memory), then rightsizes storage and disk throughput — executing each as a distinct step in the workflow, within your defined maintenance window. The result is a complete optimization of the database infrastructure without manual approvals at each stage.
Current support is strongest for Postgres and MySQL. Oracle and additional database engines will benefit from an expanded metric set in a future release, including deeper signals like cache utilization, connection counts, query type distribution, and I/O throughput, enabling more precise rightsizing recommendations across a broader range of workloads.
Built for Scale
The impetus for this launch came directly from customer demand. At a certain scale, Copilot's approval-based model stops being practical — whether that's hundreds of VMs spread across regions or thousands of RDS instances that need continuous rightsizing. The volume of potential optimizations simply outpaces what any team can manually review and approve.
Autopilot was built to be that system: methodical, safe, and capable of operating at the volume that modern cloud environments require, without creating a backlog of pending actions for your engineers to work through.
What's Next
This release is the foundation. For RDS in particular, our team is actively developing a richer metric set that will improve rightsizing accuracy and expand autonomous optimization coverage to additional database engines. Expect a V2 release that goes significantly deeper.
In the meantime, if you're already using Sedai and running VMs or RDS in Copilot mode, reaching full Autopilot is straightforward. The decision engine is the same — the difference is that Sedai now acts on its own recommendations, at the right time, within the guardrails you've configured.
Ready to let Sedai take the wheel on VMs and RDS?
Book a Sedai demo to speak with a technical expert.
