Large language models, while powerful, suffer from a lack of persistent memory, requiring repeated instructions. This led to the development of agentic AI, which can perform tasks autonomously rather than just offering suggestions. The author explored early GitLab AI features and then moved towards more capable agentic tools like OpenCode. The key to effectively using AI assistants is guiding them with specific engineering instincts and clear directives. An optimized AI workflow prioritizes tasks, allows for rapid context loading, and enables parallel work on multiple projects. Recurring tasks can be automated through documented procedures for consistent execution. Token efficiency is crucial, achieved by using optimized tools for data gathering instead of relying solely on the LLM. Active delegation where the AI executes tasks under human oversight is a core pattern. Vague instructions to AI are ineffective; precise, surgical directives are necessary for reliable results. To prevent AI sessions from interfering with each other, coordination primitives like worktrees and memory systems are essential. Existing solutions should be explored before building new ones; contribution is often preferable to reinvention. Proactive context injection, where relevant memories are automatically surfaced, significantly enhances AI efficiency. This shift from active to passive recall allows AI to anticipate needs and provide context without explicit prompts. Boot gates and iterative refinement are vital for improving AI performance, ensuring they pause and load critical directives.