How would you want an AI autoscaler to reason about scaling?

[hwclass]

Hey everyone 👋

I’ve been experimenting with Docker’s new AI stack — Model Runner, cagent, and MCP — to build something I call Docktor 🐳🩺, an AI-native autoscaler for Docker Compose.

Instead of static scaling rules or thresholds, Docktor uses local LLM reasoning to decide why and when to scale up or down — based on live container metrics like CPU usage.

I’d love to hear from the community:

If you had an AI autoscaler running in your Docker setup — how would you want it to think?

Some questions to spark ideas from you:

• Should it predict load before it happens (e.g., “Monday morning traffic spikes”)?
• Should it explain every decision (“Scaled up 2x due to consistent 85% CPU over 2 mins”)?
• Would you trust it to autonomously act, or prefer a “propose + approve” flow?
• What metrics matter most to you — CPU, memory, network, or user traffic?
• Should it adapt its aggressiveness based on business hours or SLA goals?

If you’ve used Kubernetes Horizontal Pod Autoscalers, Prometheus, or even custom scripts, what’s something you wish they did better?

Let’s discuss how we could make scaling intelligent, explainable, and local-first.

Feel free to drop your thoughts below 👇 🤗

📎 Context / Links
Docker Model Runner, cagent, MCP

[clemenspeters]

As mentioned in https://www.linkedin.com/feed/update/urn:li:ugcPost:7387501476232900609?commentUrn=urn%3Ali%3Acomment%3A%28ugcPost%3A7387501476232900609%2C7389598203366772736%29&dashCommentUrn=urn%3Ali%3Afsd_comment%3A%287389598203366772736%2Curn%3Ali%3AugcPost%3A7387501476232900609%29, scaling based on NATS queue backlog (count of pending messages) would be great