A support ticket starts in English. The customer replies in Spanish. The next message is a mix of both, plus a screenshot with Chinese UI copy.

Most tools treat that as three separate jobs. You paste into a translator, lose tone, then paste back into a chat that has already forgotten the first thread.

The failure is not vocabulary. It is context.

Language switching mid-conversation is the real job. The agent has to keep the same customer, the same constraint, and the same next action while the words change. Parallel threads matter for the same reason: one person is rarely doing one thing. You are drafting a reply, summarizing a call, and translating a policy at the same time.

What actually works in practice:

1. Keep one thread per customer or deal. Do not dump every language into a single dump chat.

2. Paste the original, not your recap. Recaps flatten the register and the agent copies the flattened version.

3. Ask for the output language explicitly, then ask it to preserve the original as a reference block.

4. If it misses tone, send the bad output back with one sentence of correction. Do not start a new thread.

That is the bar we built Nexus Agent against — fluent switching, multiple threads, no copy-paste relay.

If you run support, sales, or ops across languages, try it on a real thread you already lost time on. That is the only honest test.