Agentic Editing

One instruction, and every change it brings with it.

A term that is now feminine, so its article elides and the words around it agree. A figure that moves past four, so the Russian noun after it takes another case. A UK spelling that depends on whether the word is a noun or a verb. A Brazilian text made European, vocabulary and grammar together. A client's inclusive-language rule. Button labels rewritten to fit a limit. A house style that takes the superlatives out. Each is one instruction, typed as you would say it to a colleague.

It proposes. You decide. Every time.

The agent suggests an edit; you accept or reject it, segment by segment. Nothing it produces reaches your target text without you. Every route into your target text goes through that queue: it is how the feature is built.

See it in the applicationHide the screenshot

That nothing reaches the target without the translator — the translator writes the instruction and accepts each edit, and the machine only finds where it applies. This is how the feature is built, not a setting — The proposal queue with two edits accepted — word-level diffs in the target cells, the Diff toggle, and the console carrying the translator's instruction

Try it on a few, then the rest.

Ask for the change and look at the first proposals. If they are not what you meant, say so and the next ones follow your correction; when they are right, ask for the rest. The agent searches the file up to nineteen segments at a time and proposes edits ten at a time. When a change reaches further, it tells you how many segments first and asks whether you want them all at once or in batches.

It knows which segment you are on, when you are on one.

Put the cursor in a unit and the agent has its context. Away from the grid, it works from what you asked.

Protected by the same confidentiality profile as your translation.

Your instruction, the segments the agent reads and everything its searches return go to the AI model with stand-ins in place of the names your profile protects, from the same rules you set for translation. Every request the agent sends is read back before it goes, the whole conversation each time, and refused if a protected name is still in it. The proposals you review carry the real names.

Runs on an AI model you choose for it.

The agent needs an AI model that can work with tools over several turns, and you choose it in the Agent Console from those that can: a wide choice, across five providers, including open models served by hosting providers. The list is refreshed when the application starts, so a new model can appear without a new version of AV. Add the provider's key in LLM Configuration and its models are there to choose.

Ask your translation memories