Notes from inside the work.
Short dispatches on AI systems, automation, and FPV. Written while the systems are still running, not after the fact.
41 of 41 tests passed. Three of them were failing.
Four checks in two days that were structurally incapable of reporting a problem. One habit behind all of them, and the question I now ask before calling anything verified.
I let Claude Fable 5 spin up 62 agents and burn millions of tokens in one session
Anthropic dropped a new model today. I threw a full site review at it and told it to use as many agents as it needed. 62 agents, adversarial verification, 52 confirmed findings, one session.
What the operator stack actually does now
A field note on what changed after moving more daily work through Hermes: control surfaces, verification, memory, and the parts that still feel too manual.
A safer lane for AI code triage
What I kept from ClawSweeper was the product shape: local evidence, explicit proposals, stable hashes, and a human apply boundary.
AI ops on an early-access game server
A Windrose server went live on a Windows 11 VM the same hour a friend asked, then stabilised through practical AI-assisted operations.
The operator control plane is the product
Why the useful AI stack is the surface around delegation, verification, memory, and deployment. The model choice is the smaller part.
AI design tools need an apply boundary
Frontend generation is useful when it produces artifacts, evidence, and reviewable winner bundles, not silent production patches.
Local-first memory for agents is an operations problem
Persistent memory helps only when it reduces repeated steering without becoming a second, fuzzier source of truth.
Boring website ops beats heroic redesigns
Status, crawlability, deployment freshness, screenshots, and the regressions users actually feel.
I built a multi-agent AI system that actually works
A layered multi-agent system with 8 agents, local LLMs on Ollama, and Claude as orchestrator. Research, review, and QA, 24/7.