We are building AI agents like craftsmen. Who will certify them before they go to work?

In my previous post, I compared the invention of writing 5,000 years ago with the emergence of ChatGPT. My argument was that major technological revolutions don't just create new capabilities. They create entirely new professions, institutions, and economic necessities. I've been thinking about what one of those new institutions might look like. What if the next major AI industry isn't about building agents, but about certifying them? 1. We're still in the craftsmanship era Look at how AI agents are built today. Someone selects an LLM, writes prompts, connects tools, adds memory, defines workflows, runs a few tests, and deploys the result. It's impressive. But it still resembles a craftsman building a machine in his workshop. Every implementation is different. Testing methods vary. Reliability is difficult to compare. We have frameworks for building agents, but no universally accepted equivalent of a professional qualification system for deploying them. 2. Humans already solved part of this problem Think about how we prepare humans for professional responsibilities. We educate them. We examine them. We award diplomas and professional qualifications. We interview them, hire them, and give them probationary periods. We evaluate their performance, supervise their work, and sometimes revoke their authorization to practice. We don't assume that someone is professionally qualified simply because they can answer difficult questions. Yet we're increasingly building AI agents that may perform financial operations, analyze legal documents, manage infrastructure, or participate in medical workflows. And we often evaluate them primarily through benchmarks and demonstrations. Being intelligent is not the same as being professionally qualified. 3. What if AI agents had to become "Fit to Fly"? Imagine treating an AI agent like an aircraft preparing for commercial operation. Before deployment, the agent would need to demonstrate competence within a clearly defined operational scope. It could undergo: Professional examinations based on real human competency standards. Simulated missions, including unusual and dangerous situations. A supervised probationary period in real environments. Independent performance and safety evaluations. Certification for specific tasks and levels of autonomy. Continuous monitoring and recertification after significant changes. An agent qualified to review invoices wouldn't automatically be authorized to execute payments. And replacing its underlying model, tools, or configuration could require new testing. The certification would apply to the complete operational system, not merely the LLM. 4. From craftsmanship to industrialization This leads to another thought. Today, agents are often custom-built. Eventually, we may assemble them from standardized, interchangeable components, almost like Lego. Reasoning capabilities, domain knowledge, tools, memory, security, permissions, and monitoring could become reusable modules. But certifying individual components wouldn't be enough. The assembled agent would still need to prove that the complete system works reliably. That's how mature engineering disciplines operate. 5. Who becomes the Bureau Veritas of AI? This is where the economic question becomes interesting. Perhaps a new category of institution will emerge: independent organizations that examine, certify, inspect, and continuously assess professional AI agents. Not necessarily the companies building the models. Not necessarily the companies assembling the agents. But independent bodies whose certifications are recognized by employers, insurers, and eventually regulators. Their real product wouldn't be another AI model. Their product would be institutional trust. And their most valuable asset might be the accumulated evidence of an agent's professional competence, operational history, incidents, and reliability. Of course, a certificate would never guarantee that an agent cannot fail. It would establish what has been tested, under which conditions, and within which limits. 6. The bigger economic transformation Writing helped create scribes, accountants, administrators, auditors, and legal institutions. The industrial revolution created new systems of engineering standards, inspection, and certification. Perhaps autonomous AI will create its own equivalent. We may be witnessing the beginning of an economy in which non-human workers have qualifications, operational permissions, performance histories, and periodic recertification. And that raises a fascinating question: Will the most valuable companies of the next AI era be those that build artificial workers, or those that make artificial workers trustworthy enough to employ? If you were starting a business today, would you build AI agents — or the independent certification infrastructure that thousands of agent builders might eventually need?