Session: Multi-Agent Infrastructure Automation

Full write-up coming soon.

The multi-agent-lab is a 3-agent pipeline for Proxmox infrastructure automation: a Planner orchestrates intent, a Terraform sub-agent applies it, and a Validator sub-agent inspects results. When something fails, the Planner uses LLM diagnosis to form a hypothesis and retry.

Key design decisions:

• LLM calls only happen in the DIAGNOSING state — zero LLM calls on clean runs
• Hypothesis deduplication prevents circular reasoning loops
• Prometheus metrics with scenario + model labels for observability
• Supports LXC containers and VM clones via separate Terraform modules