Stop letting managed AI tools own your workflow infrastructure

David Andre details his personal agentic engineering stack spanning interfaces (BB, CMAX, Ghostty), harnesses (Pi, Cursor CLI, Hermes), and cloud infrastructure. He argues that serious multi-agent work requires moving beyond local execution to cloud agents on your own VPS, using Hostinger as an example. Ecosystem lock-in from managed providers is his primary concern.

His core workflow emphasizes skills as composable building blocks—reusable prompts that encode judgment about reviews, architecture decisions, and safety guardrails. He demonstrates how to have one agent provision dependencies on a fresh server, then compound that to install additional tools. The demonstration shows real agent-to-agent orchestration. The practical tradeoff is clear: this setup

delivers maximum control and cost efficiency (subscription stacking over API pricing) but demands significant upfront investment in skill authoring, terminal fluency, and workflow customization. His claim that VPS-based agents outperform managed cloud agents for serious work is supported by specific technical reasoning about session persistence and scalability, though remains largely anecdotal based on

personal experience. The underlying tension is between governance and velocity. His guardrails, logging, and architecture decision records (ADRs) are organizational patterns, not just technical tricks. Teams adopting these patterns need to budget for the maintenance overhead of custom skills and the learning curve of terminal-based orchestration, which he acknowledges but doesn't fully quantify.