Building an AI Agent Takes a Weekend. Governing It Takes Much Longer

Last weekend, the season finale of House of the Dragon aired for seventy-five minutes. Seizing that opportunity, I aimed to construct an agentic chatbot within that timeframe. Determined and focused, I watched the episode on one screen while having a terminal open on the other, challenging myself to complete the task before the credits rolled.

My bet paid off, and by the time the credits appeared, I had created a functional chat interface, a dashboard, tool access, and a model capable of receiving requests, determining what needed to be done, and executing it. In every sense, it was an agentic system, but it was entirely ungoverned. There were no limits on its actions, no record of its activities, and no human approval required before any action took place.

Essentially, it was a dragon that had not yet been trained—capable of anything and controlled by nothing. Initially, I underestimated the magnitude of what I was undertaking. It took longer than expected for me to realize the true nature of my creation. These creatures, not trained, had the power to call APIs, query databases, select tools, and pass information among themselves, often doing things I hadn't considered forbidding.

This realization marked the beginning of a week-long endeavor. Over the following days, I focused on the aspects that are rarely showcased in demonstrations. Logging became crucial, ensuring every action the system performed was recorded and preserved beyond a single session. Attribution became essential to differentiate between actions initiated by humans and those carried out autonomously by the model.

Explicit intervention points were introduced, allowing for human intervention if necessary, rather than the system proceeding autonomously. Boundaries were set, restricting the model's reach to specific systems, tables, operations, and outlining the consequences of requests that fell outside these boundaries. These measures were designed in line with the EU AI Act as a reference guide, although most of the system's operations fell outside the high-risk category specified by the act.

The act's comprehensive description of what a system needs to be accountable for itself proved invaluable, even if the specific requirements did not apply directly. The creation process took seventy-five minutes to build, but a week was needed to ensure the system was safe to run, and the job was far from complete. This ratio highlights the essence of the work involved in building and governing AI agents.

It is not about the impressive weekend build but rather the ongoing effort to ensure the system behaves exactly as intended. This process is often overlooked in demonstrations, which typically focus solely on the initial build. The real work lies in the week of writing rules, testing, and maintaining control over the system's behavior.

This ongoing process, which never ends, is the true cost of building an AI agent and is often forgotten in budget planning. When teams request funding to build AI agents, they usually only consider the initial build time. However, this overlooks the significant, ongoing expenses associated with governing the system. These hidden costs include logging, attribution, intervention points, and boundary enforcement—elements that must be continuously maintained and updated.

In essence, the budget for an agentic system should account for both the initial build and the ongoing maintenance, as the latter is an ongoing operational expenditure rather than a one-time project. Therefore, when planning to deploy AI agents in production, it is crucial to learn how to tame these powerful systems before unleashing them.