Agent Foreign Language Handling
Lessons from AP French guide creation in apibstudyguide (April 2026).
Key Finding
LLM agents default to English orthographic patterns. The creator agent generated 783 lines of AP French content with zero French accents — “francais” instead of “français” throughout. Students would memorize incorrect spellings from a language exam prep guide.
Root Cause
No target-language awareness in the multi-agent-architecture pipeline. The creator, validator, and reviewer all treated content as English regardless of subject.
Fix Pattern: Research → Configure → Validate → Regress-Test
- Creator researches the target language’s orthography before writing (accents, scripts, directionality, punctuation)
- Creator documents language rules in workflow-config.json during the research step
- Validator checks content against language-specific rules (missing accents, common misspellings)
- Regression test prevents recurrence (zero-accent check for configured languages)
Critical Insight
Not all foreign languages need special handling — English, Malay, Indonesian have no diacritics. The creator must research, not assume. Language-specific validation only runs for subjects with explicitly configured rules.
Broader Pattern
This is an instance of agent domain blindness — applying defaults from one context to another without adaptation. Same pattern caused:
- Orphaned guides (homepage not updated for new subjects)
- Broken commit format (tkt close syntax wrong)
- “Coming soon” placeholders shipped to production
The fix is always: research → configure → validate → test.
See also: multi-agent-architecture, apibstudyguide, synthesis-project-patterns