notes from a self-hosted fleet.

Why the msp-tier3-toolkit scripts exist — the actual math on what an unscoped outage costs a small business, and why tier-3 runbooks pay for themselves the first time they're used.

What actually stops an autonomous agent from doing something dumb — money and outbound traffic clamped by default, and a standby box with no stake in the outcome.

A 30B mixture-of-experts model outproduces a 27B dense model by 9x on the same Arc B60 — the followup that explains why, down to what actually moves through VRAM per token.

Why running an autonomous AI fleet safely is a security operations problem first and an AI problem second — and what two decades of the former actually buys you.

The real math on running your own AI fleet vs. an API — a $500 card, a real electricity bill, and the crossover point where self-hosting wins.

A persistent, shared, self-correcting memory that every AI agent and tool reads from and writes to — so nothing gets re-explained twice. The design decisions that make it work: supersede-never-delete, write the mechanism not the answer, dedupe before you write, and automatic write-back. Past 130,000 memories, live.

Serious local inference doesn't need a 4090. Here's the exact stack — llama.cpp, Vulkan, IPEX-LLM — the driver and VRAM gotchas that actually bite, quantization tradeoffs, and the production finding: a 30B MoE at ~40 tok/s beats a dense 27B at 4.4 on the same card.