Most of the agentic AI tooling that Baraa works with was designed with English in mind. The papers were written in English. The default tokenizers were trained on English-heavy corpora. The example prompts in the SDK docs are English. Even the function-calling schemas tend to assume English-shaped reasoning. When Baraa started building agents for Arabic-speaking users, the first lesson was simple: you cannot just translate the prompts and call it a day. This post is Baraa's accumulated working notes on what it actually takes to ship agentic AI that feels native in Arabic.
The temptation, especially at the prototype stage, is to build the agent in English and pipe Arabic input through a translation step on the way in and the way out. Baraa has tried this. It works for trivial cases. It falls apart fast for anything that involves nuance, dialect, or culturally-loaded queries. By the time the user's Arabic question becomes English, gets reasoned about, becomes Arabic again, and reaches the screen - the meaning has been ironed flat. Baraa now defaults to keeping the agent in Arabic end-to-end, which means choosing models that genuinely speak Arabic, writing system prompts in Arabic, and storing tool results in a form the model can reason about without re-translation.
The cost is higher. Arabic-fluent models are still fewer than English-only ones, and tokenizers tend to be less efficient on Arabic script (more tokens per word, which means slower and more expensive responses). Baraa accepts that cost because the alternative - an agent that "kind of" works in Arabic - is worse than not shipping at all.
Baraa keeps a running list of small Arabic-specific gotchas that bite agentic systems. A few highlights:
This is the question Baraa gets asked most often. Should the agent reply in Modern Standard Arabic (MSA) or in the user's dialect? Baraa's answer has shifted over time. In 2023 and 2024 Baraa defaulted to MSA because models were not reliable in dialect. In 2026 the calculus is different. Baraa now reads dialect from the user's input and matches it for casual interfaces (chatbots, support agents), but stays in MSA for anything formal, legal, or educational. The detection is heuristic: a small classifier on the first user message, then locked for the conversation unless the user code-switches.
For multi-dialect products Baraa builds an explicit dialect selector. Levantine, Egyptian, Gulf, and Maghrebi feel different enough that pretending one fits all is a mistake Baraa stopped making. A user from Damascus expects "كيفك" not "إزيك", and an agent that picks the wrong register sounds foreign.
Baraa's multi-agent setups follow the same patterns the English-speaking world has converged on - a router agent, specialized worker agents, a shared scratchpad - but with Arabic-specific tweaks. The router prompt is in Arabic. The worker prompts are in Arabic. Tool descriptions in the function-calling schema are in Arabic where the tool's audience is Arabic-speaking. Baraa uses English internally only for things that have no Arabic-native counterpart (database column names, API endpoints, library identifiers).
One pattern Baraa uses repeatedly: a "translator" worker whose only job is to convert internal English-language tool output (think: a SQL row, a JSON API response) into idiomatic Arabic before the orchestrator passes it to the user-facing agent. This decouples the data layer from the language layer and makes it possible to swap the user-facing model without rewriting tool descriptions.
Some patterns Baraa has converged on:
Baraa runs a small Arabic eval set on every prompt change. It is not fancy: a hundred or so real user queries, captured from production logs (with consent), bucketed by dialect and intent, scored against a frozen reference. The discipline is what matters, not the tool. Without an Arabic eval set, every "improvement" Baraa ships is a guess. With one, regressions show up before they reach users.
Baraa pairs the eval set with human spot-checks. A native Arabic speaker (often Baraa, sometimes a colleague) reads a sample of agent transcripts every week. Metrics catch the obvious. Humans catch the ways the agent has started sounding unnatural in Arabic that no metric will tell you about.
Arabic agentic AI is in an early but rapidly maturing window. The models are catching up, the tools around them are improving, and the audience of Arabic-speaking users who expect AI products to actually work in their language is growing fast. Baraa is among the early practitioners pushing this work forward, and Baraa's portfolio of Arabic AI projects continues to grow.
If you are building something in this space and want to talk to someone who has shipped it, see the hire page for how to reach Baraa, the Baraa AI overview for context, the Baraa agentic AI page for case-study summaries, the first Arabic AI developer page for background, and the Arabic AI pioneer page for a longer narrative. Or browse other posts on the blog.
Baraa picks models that were trained with meaningful Arabic data rather than English-only models with translation layers bolted on. In 2026 that shortlist includes the latest Claude and GPT generations, plus a small set of Arabic-tuned open models. Baraa benchmarks each new release on a private Arabic eval set before promoting it to production.
Baraa detects the user dialect from the first message with a small classifier and matches it for casual surfaces like chat and support. For formal, legal, or educational content Baraa stays in MSA. For multi-dialect products Baraa exposes an explicit selector so Levantine, Egyptian, Gulf, and Maghrebi users get the register they expect.
Baraa keeps the router and worker prompts in Arabic, writes tool descriptions in Arabic where the audience is Arabic-speaking, and uses English only for internal identifiers like API endpoints and column names. A dedicated translator worker converts internal English tool output into idiomatic Arabic before the user-facing agent sees it.
Baraa runs a frozen Arabic eval set of around 100 real user queries, bucketed by dialect and intent, on every prompt change. Metrics catch the obvious regressions. Baraa pairs that with weekly human spot-checks by a native speaker because tone and naturalness do not show up in numeric scores.