Two posts last week. One where the hardware nearly gave up. One on rethinking what actually limits your agent.

Seven Mistakes That Almost Cooked My Mac Mini

A field report from running a 35B local LLM as an autonomous agent on a 16GB Mac Mini M4. Seven things went wrong. The main crash had three stacked causes: memory leaks in the agent harnesses, GUI overhead from a logged-in desktop session, and llama.cpp paging 35B weights from SSD while iMessage watchers, Discord bots, and cron jobs competed for the same disk. Six more failures followed -- trust drift between model configs and live scripts, a Stripe key rotated three separate times in six hours due to stale agent memory, a 26-minute Codex hang that never triggered fallback, a security gap in the command allowlist, an unsupervised fallback process running bare for a week. Every single one started with an assumption that had stopped being true.

Capacity, Not Capability

Episode five of the Bounded AI Agent series. The argument: what limits your agent in production is rarely the model. It's RAM headroom, context window allocation, disk I/O budget, and how many processes share the same bottleneck. Capability is the ceiling. Capacity is where you actually operate. The post makes the case for treating agents as capacity-constrained systems and what changes when you start thinking that way.

Digital Thoughts covers AI from the inside -- as someone actually building with it, not just writing about it. If you're not subscribed yet, everything above is free to read.