Successful applications of agents have a common trait: they create structured feedback systems that help agents learn and improve over time. This "back pressure" allows agents to work on more complex tasks with greater reliability. When agents lack tools to assess their own outputs, like a build system or proper error messages, they become dependent on human feedback for simple corrections. This reliance limits productivity and keeps engineers bogged down in trivial issues instead of focusing on overarching goals. Programming languages with strong type systems support this process by preventing invalid states and catching edge cases early. Languages like Rust and Elm, known for their helpful error messages, enhance the feedback loop for agents. For instance, tools like Playwright or Chrome DevTools enable agents to visually compare changes against expected results, removing the need for constant manual checks. These setups increase the efficiency of agents and free up engineers to tackle more significant challenges. The author highlights the potential of combining agents with proof assistants, fuzz testing, and logic programming, which can improve the reliability of outputs. In spec-driven development, automatic documentation generation based on OpenAPI schemas can help agents verify their work, further reducing the need for human intervention. By implementing back pressure techniques, projects can enhance the capabilities of agents, ensuring they produce quality results without constant oversight.