Quick answerThere is no industry wide license comparable to a professional engineering stamp, so most companies need to build their own internal competency standard rather than wait for one. That standard should cover four areas at minimum: the mechanics of the underlying model and its failure modes, prompt and context design specific to your production stack, evaluation methodology so a builder can tell whether a change actually improved behavior, and the domain and policy knowledge specific to what the agent is allowed to say. Treat it as a role specific onboarding and recertification program, not a one time class.
Why this is a different question from reviewer qualifications
It is tempting to assume that whatever training you already require for the humans who review the agent's output after the fact also covers the people who build it. It does not. A reviewer's job is to judge whether a specific answer was correct or appropriate, which is a downstream, output focused skill. A builder's job is to shape the system that produces every answer at once, which requires understanding the mechanics that generate those answers in the first place. A company can have a rigorous review qualification program and still have undertrained builders, because the two roles exercise different judgment at different points in the pipeline.
The model mechanics layer
A builder needs a working understanding of how the specific models in production actually fail, not a generic AI literacy course. That means knowing the concrete difference between a hallucination driven by a knowledge gap and one driven by an ambiguous prompt, understanding how context window pressure degrades instruction following in long conversations, and knowing the particular quirks of whichever provider and model version is currently live. This layer goes stale as models change, which is the strongest argument for recertification on a cadence rather than a single onboarding module.
The evaluation methodology layer
The single most useful skill a builder can have is the ability to tell, with evidence rather than intuition, whether a prompt or architecture change made things better or worse. That means fluency with your golden evaluation dataset and with reading eval results correctly, not just running them. Builders who ship changes based on a handful of manual spot checks rather than a structured eval run are the most common source of silent regressions, and this is a trainable, testable skill, not just experience that accumulates on its own.
From the team
We build production AI systems for startups.
LLM pipelines, RAG, and agent workflows that hold up under real traffic — not just in the demo.
The domain and policy layer
Nobody should be authoring prompts that determine what the agent can promise a customer, refund, or commit to without first passing a competency check on the actual policies those prompts encode. This overlaps heavily with the work already done during pre build discovery and requirements gathering, and the standard should require a builder to demonstrate they understand the requirements document, not just that they read it once during onboarding.
Where prompt engineering skill fits and where it does not
Prompt authoring is a real, specific skill, related to but distinct from the broader question of prompt engineering versus fine tuning as an architectural choice. A certification standard for builders should test whether someone can write a prompt that reliably produces the intended behavior across edge cases, not just whether they know the vocabulary. Pairing a junior builder with a senior one for their first production prompt changes is a cheap, effective control while the formal standard is still being built out, and most teams end up running both in parallel rather than replacing one with the other.
FAQ
Is there an external certification worth requiring instead of building an internal one?
Model provider certifications exist and are worth encouraging, but they teach the provider's tooling, not your production system, your domain policies, or your specific failure history. Treat external certifications as a useful supplement to internal training, not a substitute for it.
Who reviews whether a builder's judgment calls are still sound after they are certified?
That is a separate, ongoing question from initial certification, closer to the territory of who is qualified to review the agent's decisions after launch, which deserves its own standing program rather than being folded into a one time builder credential.