User guide · Part 4 bAIbel AV
English
Draft

Agentic Editing Mode

Sometimes you need to make the same kind of change across a whole translation: switch a preferred term, change how a product is named, or shift the level of formality. A find-and-replace can do the blunt version of this. Agentic Editing Mode does the intelligent version. You describe the change in plain language, an AI agent finds the segments and proposes edits, and nothing reaches your file until you have validated it.

What Agentic Editing Mode is

Agentic Editing Mode lives in the Agent Console. Open the console from its toolbar button or with Ctrl+`, then set the mode dropdown to Agentic Translation Editing. You pick a model, then type an instruction such as “find every place we wrote colour and change it to color, keeping the grammar correct”.

The agent works through your document, searches for the relevant segments, and proposes a rewrite for each one. Every proposal waits for your verdict.

Figure 1. The Agent Console in Agentic Translation Editing mode. Choose a model, then type your instruction in the Ask the agent… box and press Send.

The Automate–Validate principle

The mode is built on a simple safeguard we call Automate–Validate (AV). The agent automates the tedious part — finding the segments and drafting each edit. You validate the result — you review each proposal and decide. Only the edits you validate are written to the finalized file. It is an expert-in-the-loop design: the agent does the labour, the translator keeps the judgement.

Only validated changes ripple through

A proposed edit is never silently applied. It sits in a review queue as “pending” until you accept it. When you choose Accept All & Confirm, only the accepted edits are committed to disk. Anything you reject simply disappears.

This is the opposite of a find-and-replace, which changes your file the instant you click. With Automate–Validate, the agent proposes and you dispose.

Why it matters: your authentic voice

“AV” carries a second meaning — Authentic Voice. Because nothing is applied without your approval, the agent never quietly overwrites your work. It does the searching and drafting; the final translation stays yours. The voice in the delivered file is the translator’s, not the machine’s.

Where it beats Replace All — even regex

A find-and-replace matches characters. It does not understand language. For many real translation changes, that is not enough — and the gap is widest in languages with rich grammar. A blind replacement leaves broken agreement behind; the agent fixes the whole sentence so it still reads correctly.

The examples below all start from the same kind of request: “change this term throughout”. Watch what a plain Replace All gets wrong.

German — the article must follow the noun’s gender

Change the masculine term der Laptop to the feminine die mobile Workstation.

ApproachResult
Replace Allder mobile Workstation — wrong. The article der was left untouched and no longer agrees.
The agentdie mobile Workstation — correct. It changed the article and adjusts the surrounding words to match the new gender.

French — agreement and elision ripple outward

Replace the feminine la solution with the masculine le logiciel, in a sentence like « la solution est conçue pour… ».

ApproachResult
Replace Allla logiciel est conçue — wrong twice. Wrong article, and the past participle still agrees with the old feminine noun.
The agentle logiciel est conçu — correct. Article, adjectives, and participles all follow the change. It also handles elision, turning l’application into le système rather than l’système.

Spanish — contractions form silently

Replace the feminine la herramienta with the masculine el programa.

ApproachResult
Replace Allde la programa — wrong. Spanish requires the contraction del, which a string replace never produces.
The agentdel programa — correct. It forms de + el = del and a + el = al wherever they apply.

Russian — one term, many case endings

A noun in Russian changes its ending depending on its role in the sentence. The same term appears as программа, программы, программе, программу, and more.

ApproachResult
Replace All (even regex)Either it misses the inflected forms, or a stem pattern mangles the endings and produces non-words.
The agentIt recognises every inflected form and writes the correct case ending for each occurrence.

Beyond single terms — changing register

Ask the agent to move a whole document from informal to formal address — French tu to vous, or German du to Sie. This is impossible for Replace All, because it changes verb endings, pronouns, and possessives throughout every sentence. The agent treats it as one instruction and rewrites each affected segment correctly.

The pattern behind every example

Replace All changes a string. The agent changes a meaning and repairs the grammar around it. Whenever a change has to ripple through agreement, case, contraction, or register, the agent is the right tool.

It also avoids the wrong matches

Replace All has the opposite failing too: it changes things it should not. A term that should only change in one sense — a brand name that is also an everyday word, say — gets replaced everywhere by a blind search. Because the agent reads each segment in context, it edits only the occurrences that genuinely match your intent, and leaves the rest alone.

Running an editing session

  1. Open the Agent Console and select Agentic Translation Editing.
  2. Pick a model from the model picker.
  3. Type your instruction in the Ask the agent… box and press Send. Use plain language; describe the change and any conditions.
  4. Watch the agent work. It shows its steps as cards in the conversation — searching for segments, then proposing edits.
  5. Review the proposals in the Agentic Editing Mode banner and grid (next section).
  6. Commit the ones you want with Accept All & Confirm.
Figure 2. The agent’s activity. Each step appears as a card — it searches the bilingual segments, then proposes edits for the ones that match.

The review banner

When the agent proposes edits, the Agentic Editing Mode banner appears above the grid. It counts how many edits are pending (awaiting your verdict) and how many you have accepted so far.

The banner gives you the controls that decide what reaches your file:

ControlWhat it does
DiffShows a word-level comparison of the original target and the proposed one, so the change is easy to see.
Accept All & ConfirmCommits all pending and accepted edits to disk. This is the only action that changes your file.
Reject AllDiscards the pending edits. Anything you already accepted per row still commits.
Clear SelectionEmpties the review grid without discarding anything, ready for the next search.
Exit (×)Leaves Agentic Editing Mode. It asks first if edits are still waiting.
Figure 3. The Agentic Editing Mode banner. It shows the pending and accepted counts and the commit controls. Nothing reaches the file until you choose Accept All & Confirm.

Judging each edit

Turn on Diff to see exactly what changed in each segment. Then judge them one by one, or in bulk:

Figure 4. A proposed edit with the diff shown. The word-level highlight makes a grammar-aware change — here the article changes with the noun — easy to verify before you accept.
Figure 5. Per-segment controls. A accepts and R rejects, and the information icon reveals the agent’s reasoning.

Privacy and cost while the agent works

The privacy promise from guide 1 holds here too. When the agent needs to compare meaning across segments, your text is passed through the project’s confidentiality pipeline before it leaves your machine. A badge tells you whether the content is masked or being sent as RAW source, so you can stop and configure protection first if you need to.

The console also shows a running cost for the session — the tokens sent and received, and the amount in dollars — so there are no surprises.

Figure 6. The masking badge and the session cost meter. You always know whether content is protected and what the session is costing.

When a plain Replace All is still the right tool

Agentic Editing Mode is not for every change. If you simply need to swap one fixed string for another — a product code, a URL, a date format — and there is no grammar to repair and no risk of wrong matches, an ordinary find-and-replace is faster and cheaper. Reach for the agent when the change has to understand the language: when it must ripple through agreement, case, contraction, or register, or when it must tell apart occurrences that look the same but mean different things.

Terminology used in this guide

Agent Console
The panel where the AI agent runs, opened from the toolbar or with Ctrl+`.
Agentic Editing Mode
The mode in which the agent proposes edits across your translation for you to review.
Automate–Validate (AV)
The principle that the agent automates finding and drafting edits, while you validate them; only validated edits reach the file. AV also stands for Authentic Voice: the delivered translation stays the translator’s own.
Pending
A proposed edit that is waiting for your verdict.
Accepted
An edit you have approved per row; it commits when you confirm.
Accept All & Confirm
The action that commits all pending and accepted edits to the file. The only step that changes your file.
Diff
A word-level comparison of the original and proposed target text.