Linux Is 10x Faster for AI Agents—Time to Ditch the MacBook

The biggest concrete signal is file system performance: APFS on Macs drags git clean and pnpm install to 30x slower than Linux ext4—crippling agent loops that rely on frequent worktrees and subagents. This isn't a minor annoyance; it directly limits how many agents a developer can run concurrently.

Implication: A maxed-out M5 Max MacBook Pro becomes unusable under Codex subagent load, with macOS’s CIS policyd process spiking CPU. The fix is offloading work to local Linux machines accessed over Tailscale and remote IDE tools, freeing the laptop from compute and heat.

Tradeoffs include managing a fleet (orchestration overhead), losing Apple’s unified memory for local models, and needing network KVMs for full remote control. However, the subsidized token plans from Codex and Claude deliver disproportionate value when run on personal hardware, not cloud IDEs.

For engineering leaders, this shifts infrastructure into the developer workflow. Provisioning purpose-built Linux boxes, treating agent orchestration as a discipline, and rethinking hardware purchasing become critical. Cloud-native dev environments aren’t ready; self-hosted fleets offer more control, speed, and cost efficiency.