Can your laptop defend itself from its own agents?

Hypothesis

Running many AI agents locally means occasionally one of them — or their tooling — eats the machine. A small watchdog daemon should detect runaway resource use and intervene before the laptop becomes unusable, without a human noticing anything except a notification.

What we built

Tripwire: a lightweight daemon that watches the Mac's health and steps in when things go wrong — runaway processes, memory-pressure spirals, background indexers gone rogue. It escalates gently (notify → throttle → kill the offender), keeps a journal of what it did, and pings the phone so interventions are never silent.

Learnings

• CPU load is the wrong trigger: a busy machine is often fine. Memory pressure is the signal that predicts an unusable laptop — the redesign that mattered was "load never kills, memory pressure does".
• A watchdog needs a single owner per resource it manages; two daemons toggling the same system flag fight each other forever.
• Guardrails for agents are the same discipline at every scale — the laptop daemon and the server approval gate are one idea.

Log

• 2026-07-10 — Memory-pressure redesign: load alone never kills; pressure-based escalation live.
• 2026-06-15 — First watchdog with notify/kill escalation.