Posts

Unlike 2–3 years ago, a prompt can now be considered an executable program. This is because an agent can access local files and execute simple programs on its own. Furthermore, agents can fetch prompts on their own based on the tasks they are given. This means we cannot statically determine which prompt will be executed. For instance, if we instruct an agent to review code, the agent may fetch the ‘how-to-review-code.

My first project was Synstiller, which aimed to build an interpretable coding model. Over nearly a year and a half, I invested substantial effort to solve the problem, but I did not achieve clear results. However, knowing I did my best allowed me to let go with confidence, teaching me that doing my best is the key to moving on without regret. After that, I worked on a project called VeriSafe Agent, which verifies an agent’s behavior from a program verification perspective.

AI capabilities are improving exponentially 1, with recent benchmarks showing models outperforming human developers. Yet, we still struggle to construct large-scale, high-quality software using AI. Attributing this solely to a lack of specific training (e.g., DevOps) overlooks the fundamental issue. The core challenge is alignment: ensuring the AI executes the user’s true intent. As task duration increases, alignment becomes increasingly fragile. A microscopic initial deviation results in a massive miss over a long distance.