Cole Wisniewski ← all projects

Biscuit

In progress

A test rig for my AI agents to repair broken k3s VMs and check what changed.

Stack

Infra
Proxmox · k3s
Runner
Python · SSH
Evidence
eBPF

What it is

I built Biscuit to see what an AI agent really changes when it works on Kubernetes. It gives a coding agent a broken k3s task on a disposable VM. I compare the final cluster state with an independent check and keep the run evidence.

Happy biscuit on a peach background

How it works

I roll a Proxmox VM back to a clean snapshot before each trial. The agent runs commands over SSH while the runner records what it did and what changed. A separate verifier checks the task result. I save the trajectory, audit evidence and cost, then reset the VM.

Flow

  1. Clean Proxmox snapshot
  2. Broken k3s task
  3. Agent commands over SSH
  4. Independent verification
  5. Evidence and reset

The problem

The first full receipt path slowed short commands because it gathered cluster inventories around each command. I saw this in paired timing runs and traced much of the cost to those inventories. A hooks only run was still slow, so the tracer alone did not explain it. I moved to persistent collection and returned bounded receipts with command output. I kept failed attempts in the record.