Language nuances need explicit handling 

On the technical side, translation systems and multilingual LLMs both have limits when handling language-specific nuances. On top of that, brand and style add another layer that further complicates this.

Translation alone is not enough: It is tempting to think that we can have the agent produce its output in English and translate it into the target language afterward. The gender agreement example above shows why this falls short. An English sentence about a candidate can stay grammatically gender-neutral, but the same content in French or German requires matching grammatical gender across articles, nouns, adjectives, and titles. Translation cannot recover information the English never encoded. The agent has to generate output natively in the target language for these patterns to come out right.

Multilingual LLMs alone are not enough: Modern LLMs have strong multilingual capabilities, but the quality of their output is shaped by the prompt, not just the model. Effective prompts need concrete examples and explicit style guidance, and quality drops sharply without them. These examples and guidance are inherently language-specific. An English example teaches the model English sentence structure and vocabulary, and the model carries those patterns into other languages it generates. Even the noun capitalization rule in German, which the model knows in theory, has to be reinforced in the prompt to be reliably applied. Larger multilingual models can handle some of these nuances better on their own, but at serving costs that are impractical when each recruiter requests fans out into many model calls across the system. The LLM landscape itself is evolving rapidly, so any approach tied to a specific model's behavior is inherently fragile.

Brand and style guidelines need explicit local rules on top of the linguistic ones: Correct grammar is only the baseline, and professional norms still differ across markets. The form of address to use, the salutations and closings that are appropriate, the right level of directness: these are brand and policy decisions, not grammar, and have to be made explicitly for each market. The agent has to follow LinkedIn's brand voice in each market while adhering to local professional conventions.  

Measurement is bottlenecked by linguist annotation

On the process side, measurement requires expert linguist judgment, and that expertise is scarce across the languages and sub-agents we need to support.

Only linguists can judge language quality: Every prompt change required expert review of the outputs to verify it was an improvement. There was no automated way to tell whether the output read as native in a target language, so each iteration depended on linguist annotation. A single prompt could take weeks of back-and-forth because each round of changes needed expert review before the next round could begin.

Linguist supply does not scale across languages and sub-agents: Qualified linguists for each target language are a finite resource. The work of evaluating outputs multiplied across the many prompts in each sub-agent, the supervisor, and a roadmap of more than ten languages. The demand for linguist time far outpaced what could be delivered, and this dependency could never keep pace with the product.

Scale, surface area, and constant iteration

The work of getting language right multiplies across many surfaces. Hiring Assistant is powered by a large and growing number of prompts spread across the supervisor and its sub-agents. A single recruiter request can flow through many sub-agents, trigger several planning passes, and produce several intermediate reasoning steps, and every one of these is language-sensitive. Every pattern has to be handled at every surface, in every supported language.

Internationalization is also not a one-time job. The set of prompts is not static. New capabilities are added, sub-agents are extended, and existing prompts evolve as the product improves. Every change to an English prompt needs to propagate across every supported language. The cost compounds with every new language, every new sub-agent, and every prompt update.

Our solution

The breakthrough came when we realized the linguistic rules we kept rediscovering weren't tied to individual sub-agents. They were properties of the language itself. Once we moved them into a shared framework, the rest of the system fell into place: a structured rubric framework, automated prompt transformation, an adaptive model strategy, and a quality loop with language judges and expert feedback. 

Abstracting language nuances into a reusable framework

Language nuances are captured once in a shared rubric framework that serves as the single source of truth for language quality across the system. The rubrics are then applied through two complementary paths depending on how each sub-agent runs: at the prompt layer via an automated transformation pipeline for sub-agents using instruction-following models, and at the model layer via per-language adapters for sub-agents using smaller, cost-efficient models. Both paths draw from the same rubrics, so language expertise is captured once and reused everywhere.

The rubric framework captures the rules: We encoded the abstraction into a framework covering dimensions such as language purity, tone and register, orthography, cultural adaptation, and others. Each rubric captures linguist-reviewed, language-specific rules and serves as the shared source of truth for prompt transformation, model adaptation, and quality evaluation. The language expertise is captured once and reused across every sub-agent, so when a new capability is added, it picks up the full set of linguistic rules for every supported language with no extra work. Discovery of the rubrics is also partially automated through an LLM that proposes language-specific patterns and rules drawing on internal style guides and language conventions. The linguist team validates these proposals and adds new rules or updates existing ones based on their own expertise. New patterns also can enter the rubric over time as the language judge surfaces new patterns and as recruiter feedback identifies edge cases.