Delegate a localization change to an agent
Give an assistant a concrete outcome, bounded authority, and a result that can be independently inspected.
An agent can inspect content, prepare guidance, and operate supported localization tools. A useful assignment tells it what result is needed, what may change, and what evidence should come back. “Improve our localization” leaves those decisions unstated.
Choose a task small enough to review and recover: a selected translation, one new locale, one reported defect, or one proposed glossary entry. Connect through MCP setup and inspect the actual tool inventory and organization scope before assigning writes.
Match tools and authority to the task #
MCP tools operate on authorized organization records. Use the CLI workflow to translate files in a checkout.
Permissions provide the security boundary. A prompt requesting observation does not remove a connected credential's write authority. Use appropriately scoped access alongside explicit instructions.
An illustrative assignment could read:
Inspect the selected messages at the recorded source revision for target locale
de. Read their catalog and engine guidance. Show the proposed file or setting changes for approval. Then change only the selected messages, retrieve the result, and check variables and meaning. Keep unrelated translations and their baseline intact. Report any work awaiting review or failing to complete. Ask before changing shared guidance or publishing.
Include the relevant language-addition, model-comparison, or diagnostic method in the assignment.
For configuration, inspect before creating #
Specify a product surface, target locale, and source material authorized for the service. Keep private context private; do not turn it into a public URL for convenience. Ask the agent to inspect existing localization engines before proposing a new one.
Fill an engine from your website. Your assistant can create a new engine and fill its glossary, ruleset, and brand voice, or fill an existing engine: only the missing resource types, or a new glossary or ruleset alongside the current ones. The source is your website or text you provide.
Your assistant can check progress while the job runs and report each part as completed, skipped, or failed when it finishes. It can retry failed parts the engine still has no resource for. See Fill from your website.
If a checkout will use it, link that exact ID through the CLI workflow. Inspect a representative translation and counterexample before accepting the configuration.
For translation, preserve the request and result #
Use translation jobs for application data or uploaded files. Supply the engine, source locale, target locales, and document keys. Save the returned job ID before uploading missing files and starting the job.
Supply context and hints for ambiguous text, and state the intended source language explicitly. Poll the job, inspect failures and review waits, and download each completed target. Submission, an open review, and a downloaded translation are different outcomes.
For local catalogs, inspect the source and accepted translations, then use the CLI workflow with the linked engine to translate them. Use CLI or CI when the lasting result should be a repeatable repository workflow. Retain the source revision, run state, and output the next operator will need.
For review, name the decision being delegated #
An ad-hoc check of pasted source and target can identify suspected problems. It does not approve a pending human-review record. A measurement answer, suggestion application, and human decision change different things.
For a supported pending review, inspect the exact content and locale, relevant guidance, and unresolved work; act only with the required reviewer authority. Accepting the target and changing reusable guidance remain separate.
Require evidence of the completed result #
Ask the agent to return the source revision, changed files or record IDs, saved settings, retrieved output, checks performed, and remaining failures or access needs. Inspect the selected work before expanding the task.
Keep merge and publication outside the assignment unless expressly authorized. If a missing capability stops the task, retain the completed inspection and proposed change so a person can continue from the evidence already collected.