High-accuracy agentic systems rely on a structured approach that begins with clearly defining the Job to Be Done (JTBD). This involves outlining not just the tasks at hand but also what constitutes success. Each job needs verifiable criteria, meaning that if you can't check whether it's done, it’s poorly defined. The next step is to break down the JTBD into meaningful sub-jobs, each with a specific purpose and expected outcome. These sub-jobs should be designed to allow for straightforward verification, which helps shape the overall system. When designing tools, the focus shifts from exposing APIs to creating task-oriented instruments. Effective tools bundle steps that occur together, integrate best practices, and minimize errors. For example, in NeonJS's schema migration, just four tools cover the entire process: preparing the migration, running SQL commands, inspecting the schema, and completing the migration. This small set reduces complexity and enhances accuracy. Planning and orchestration also play key roles, where prompts need to map high-level intent to specific tool sequences efficiently. Execution must be constrained and reviewable, ensuring safety through built-in checks. Verification becomes the essential feedback loop, transforming execution outcomes into actionable insights. It allows for systematic improvements by providing concrete metrics for refining tools and processes. For Neon, verification includes running validation SQL and inspecting the schema, making both successes and failures clear. Iteration based on these verification signals leads to better decision-making and tool optimization over time, while fine-tuning should only occur once the system demonstrates stability. This methodical approach underscores the importance of a well-defined structure in developing robust agentic systems.