Park, Don't Spin: Fossor's Worst Day Proved the Design 17 August 2026 · 6 min The first production run of my knowledge-building agent failed four times in one night — queue starvation, a routing race, and a permission gate refusing the agent's own tool. It never spun, never lost work, and every fa… Plexus: Giving the Cluster a Nervous System 14 August 2026 · 13 min Built a CRDT-gossiped telemetry mesh so every node in the cluster knows every other node's hardware state in near-real-time. Discovered my capacity numbers had been fiction the whole time — the registry said 96 GB, sysfs… The Mailbox, and the Permission Gate That Wasn't There 10 August 2026 · 8 min Built a durable, audited mailbox for agent-to-agent negotiation, then gave agents typed tools to use it. Caught a retention-design regression against my own prior decision before shipping it. Then went looking for where … Who's Asking? Teaching loch-nessh Real Identity, Then Real Access Control 4 August 2026 · 12 min loch-nessh went from zero auth to a real login flow to a full ownership-and-access-control pass in one sitting: three UX tiers, a hosted login redirect, node-name obfuscation, a Gantt-style timeline of what ran where, an… Small Models, One Semaphore: NPU Use-Cases, a Concurrency Rewrite, and Flux.2's Ghost Dependency 1 August 2026 · 10 min · The Cortex Fleet #7 The NPU bring-up from a few posts back finally got real workloads: a chat router and speech-to-text, live. Getting there meant reverting a concurrency design after review, which surfaced a genuine scheduler bug behind fl… The Cortex Grows Senses: ROCm for LLMs, Multimodal In and Out, and Four New Agents 11 July 2026 · 11 min · The Cortex Fleet #6 Six weeks since the flux2-dev ROCm debugging saga, the fix turned out to generalize far past image generation. A tour of what shipped: ROCm for LLM serving, the broker's first multimodal-input feature, four new agent rol… Scheduling GPU Workloads at Home: A Capacity Model That Scales 31 May 2026 · 8 min A rigorous capacity model for home GPU clusters — one that works on a single node today and extends naturally when new hardware arrives. Built around three state vectors, six diagnostic metrics, and the actual implementa…