Workflow guides · · Frank Wei
Agent or Standard production: how to choose
Standard production drafts directly; Agent production analyzes the full document, plans terminology and register, then translates with consistency QA. Learn when to use each method.
A first draft can look acceptable sentence by sentence and still fail in review. A company name changes on page 47, dates follow two conventions, and repeated headings use different wording. No single sentence is the whole problem; the document is inconsistent, and a reviewer must correct it by hand.
To reduce that work, BabelBee offers two production methods: Standard production and Agent production.
These are not new plans or quality tiers. AI Fast, AI Reviewed, and Team Verified still determine the level of quality control. The production method determines how BabelBee creates and refines the first draft.
Free workspaces can try either production method with AI Fast or AI Reviewed within the 3,000-word monthly quota. Agent is a production method, not a paid add-on; Team Verified and reviewer invitations start on Pro.
Standard production: the faster first draft
Standard production translates directly by default, without automatic terminology analysis or an automatic whole-document consistency pass. Project terminology is available in both modes: Standard users can optionally analyze terms, import a glossary, and approve preferred translations in Manage project terminology.
That is the right default when the file is short, expressive, or mainly needed as a starting draft:
- advertisements, slogans, and campaign lines
- short copy and landing pages
- literary text, where voice matters more than a locked term list
- one-off documents that will be rewritten rather than reused
A poem, a single email, or a short marketing line usually belongs here. You still edit in the workbench, run QA, and invite reviewers. You are not giving up review. You are skipping production steps the file does not need.
When turnaround is the main constraint and the file does not depend on repeated terminology, Standard is usually the more efficient choice.
Agent production: understand the document before translating it
Agent production takes longer because it builds a document-level production plan before translating segment by segment. Context and consistency requirements shape the draft instead of being reconstructed after translation.
It works through the file in a fixed order:
- Generate statistical term candidates from repetition and context
- Run semantic document analysis to identify domain, purpose, audience, register, proper names, defined terms, and important one-off terms
- Analyze long files in chunks, then consolidate them into one translation strategy and bilingual terminology plan
- Show the plan on the upload screen: high-confidence bilingual terms with source evidence may be applied automatically; uncertain candidates stay available for a person to translate, approve, or exclude
- Translate using only applied terminology and the document strategy, then check terminology, repeated text, numbers, dates, placeholders, and omissions
- Run up to two QA repair rounds on repairable failures; unresolved findings stay flagged for people
A model-proposed source term must occur in its cited source segment or it is discarded. Existing workspace or project terminology takes priority. Explicitly approved bilingual terms are constraints; automatic suggestions are advisory. Unconfirmed or rejected candidates are not fixed translations. Changing project terms does not automatically rewrite existing translations: preview consistency changes before applying them.
If you do not click Analyze document on upload, Agent runs the analysis automatically when production starts. If the source or segmentation changes afterward, BabelBee marks the old plan as stale and rebuilds it before production.
The result is written to the AI-reviewed column. People still approve. Agent does not auto-approve rows, and it does not write the workspace master TM. Hand-edited and locked segments are left unchanged.
Choose Agent when drift will be visible and expensive:
- contracts and terms
- manuals and procedures
- reports and decks with repeated headings
- academic papers
- longer CVs
- multi-section files, or a batch in the same domain
An 80-page services agreement, a product manual, or five related reports for one client are typical cases. Agent can also identify a product or organization name that appears only once; frequency is useful evidence, not the only definition of a term. The extra time is intended to reduce later harmonization and context hunting.
Same quality tiers, different production path
The question teams ask first is whether Agent is "a better model." It is not a new tier. It is a more careful method.
Standard is a straight path from source to draft. Agent first plans the document, then spends processing time so section 12 follows terminology and register established in section 1, and so numbers, placeholders, and dates do not drift mid-document.
That additional processing takes time. If the file is short and reviewers will substantially rewrite it for voice, Standard is usually better for the schedule. If reviewers would otherwise spend hours harmonizing names, terms, and figures, Agent may reduce the total effort.
Neither method replaces a specialist on legal, medical, or certified work. Team Verified still expects a person to resolve flagged segments before export.
Where you choose the method
The control sits on the last step of upload, on the segmentation preview. Standard is the default. After selecting Agent, you can analyze the document, inspect its detected domain, register, summary, and term candidates, and revise terminology before submitting production.
If a job already ran as Standard and part of it still drifts, the workbench can re-run Agent on unapproved segments. Approved, locked, and hand-edited rows stay as they are. New output still lands in the AI-reviewed column for people to accept.
A practical way to decide on the upload screen:
- Will a reader notice if the same name or figure is translated two ways? If yes, consider Agent.
- Is the file short, one-off, or meant to be rewritten for voice? Stay on Standard.
- Unsure? Start with Standard. Unapproved segments can still go through Agent later.
What Agent does not promise
Agent is designed for consistency. It does not guarantee that every automatic suggestion is correct or that every file will come back cleaner than Standard. On high-risk content, inspect the terminology plan and keep a qualified reviewer in the release path.
Use it when the document has recurring terms, numbers, and section titles that should match. Skip it when the file is short, expressive, or you only need a first draft on the table.
The utility toolkit is a separate question. It prepares and checks language assets around a job. Agent changes how the job itself is produced. You can use both; you do not have to.
Frequently asked questions
Does Agent replace AI Fast, AI Reviewed, or Team Verified?
No. Those remain quality tiers. Agent is how the draft is produced. Team Verified still gates export.
Does Agent write to the master translation memory?
No. Output goes to the AI-reviewed column. Humans approve. Workspace master TM is not written automatically.
Does Agent force every statistical candidate into the translation?
No. Statistical candidates improve recall, but candidates without a reliable target remain pending. Only existing project terminology, high-confidence bilingual terms with source evidence, and terms explicitly approved by a person can constrain translation.
Can Agent recognize an important name that appears only once?
Yes. Frequency is not the only signal. Semantic analysis can identify product names, organizations, abbreviations, and defined terms that occur once, while requiring evidence from the source to reduce invented terminology.
Can I switch after the job has already run?
Yes, in part. The workbench can re-run Agent on unapproved, unlocked segments. It will not overwrite hand edits or locked rows.
Is Agent always the better choice for long files?
No. Length is a signal, not a rule. A long literary text may still want Standard. A shorter contract with dense defined terms may want Agent. Judge by whether names, terms, and numbers must stay consistent.