Infrastructure and Control

• Andre advocates for a self-hosted cloud-agent stack using a VPS, SSH, and a terminal-native runtime like Herder to gain persistent, isolated, and scalable environments.10:14
• He warns that vendor-managed platforms create deep ecosystem lock-in, where migration becomes prohibitively expensive once agent sessions, secrets, and environment configurations are trapped.9:29
• The most effective bootstrapping strategy is to prioritize getting one agent running, which can then handle the installation of dependencies and the configuration of subsequent agents.21:12

Operational Safety and Discipline

• He insists that dangerous "YOLO" modes or wide-permission harnesses must be shielded by global guardrails, pre-tool-call hooks, and read-only database roles to prevent irreversible production damage.34:42
• For serious projects, architectural reasoning should be recorded in dedicated /doc/adr directories, ensuring that both future humans and agents understand the rationale behind codebase choices.47:21
• When running high-concurrency agent workflows, he utilizes a "push lock" mechanism to serialize merges and deployments, avoiding the merge collisions and CI/CD congestion caused by simultaneous pushes.36:21

Model and Budget Optimization

• Andre rejects standard API pricing in favor of optimized subscription bundles, arguing that plans like the $10 Open Code Go offer superior token value due to provider subsidies.7:07
• He rejects a one-model-fits-all approach, instead matching models to tasks: Fable for architectural planning, GPT 5.6 for complex debugging, and Kimi K3 for front-end design.40:41