# Yesterday I Almost Killed My MacBook With a Swarm of AI Agents

MacBook Pro 14" M1 Pro, 16 GB RAM, ~670 processes, 3 days of uptime. The morning it wouldn't boot, I learned the hard way that a 24/7 agent swarm belongs on a server — not on your work laptop.

Author: Maxim Tolmachev
Published: 2026-06-05
Language: en
Canonical: https://maximtolmachev.com/blog/yesterday-i-almost-killed-my-macbook-with-a-swarm-of-ai-agents/
Original source: https://t.me/mtolm/84

![Yesterday I Almost Killed My MacBook With a Swarm of AI Agents](/blog-images/tg-84.webp)

The setup: a MacBook Pro 14" M1 Pro (2021), 16 GB RAM, 10 cores. On it ran Hermes — 11 launchd gateway daemons split by role, from coach and sales to cto, jobhunt, personalops and dashboard. Each one spawned Codex agents.

In parallel: ~20 Codex.app processes, 13 Claude Code sessions via cmux, the Jarvis fleet, a runner, an outbox-worker, plus the normal life of a laptop. About **670 processes** and **3 days of uptime** without a reboot.

The important nuance: this hasn't been a "clean laptop" for a long time. It runs a pile of local automation via launchd — Jarvis, Hermes, fleet workers, monitors, Telegram bots. The Mac had quietly become a home automation host, just without any server discipline.

Normally this machine opens everything instantly. This time Chrome took ~10 minutes to launch, a video took 4 minutes to open, and the login screen spun for several minutes.

Then it got worse: one morning it wouldn't power on after a night on the charger. I held it under the air conditioner in Da Nang for 10 minutes thinking *well, that's it — I killed my work computer with my home datacenter* 🫠 When it finally came back, I pinned the fans to 70% permanently.

The cause wasn't only CPU and RAM. It was the collision between Codex and **Apple Platform Security**. macOS verifies signature, notarization and integrity for software outside the App Store. For a normal user that's sane protection. For a swarm of agents constantly spawning short-lived processes, it suddenly becomes the bottleneck.

The technical detail: Codex created temporary wrapper binaries in `~/.codex/tmp/.../codex-arg0XXXXXX/`. A new random name on every run, so Gatekeeper can't reuse a cached assessment — `syspolicyd` flies to 80–95% CPU and ~700 MB RAM, `trustd` holds ~29% CPU. I left a full writeup in openai/codex #16767.

My takeaway is simple: if you want to use your computer as a server for a couple dozen agents, get a server.

A laptop can be a dev station and an orchestrator. But a 24/7 agent swarm should run where there's Linux, CPU/RAM limits, containers, logs, restart policies, cooling and a separate failure domain — especially when you also need that laptop for work.

So I'm moving Hermes/Codex and my Claude Code swarm to Hetzner and scaling the load gradually, instead of the old "oh, seems to work, let's throw 10 of them straight into prod." And it's time to upgrade the hardware — 16 GB of RAM for a life with agents already sounds like self-deception. I'm waiting for the M6 MacBook Ultra at the end of the year before deciding what to buy. Until then, I'll grow up into real server infrastructure.

I've written before about a 3.75× speedup in working with AI — but that speed has a flip side: the more agents you run, the faster "fun automation" turns into ops.

And ops isn't about which LLM is smarter. It's about who caps the memory, restarts the daemon, and doesn't leave you without a working laptop in the morning.
