Agent Domain Blindness

A failure mode in multi-agent-architecture pipelines where agents apply defaults from one domain to another without adapting to the new context.

Definition

When an LLM agent is trained or prompted in one context (e.g., English content, IB Biology) and then asked to work in a different context (e.g., French language, AP program), it carries over assumptions from the original domain. These assumptions are invisible — the agent doesn’t know what it doesn’t know.

Observed Instances (apibstudyguide)

InstanceWhat happenedWhat was blind
French accentsCreator wrote AP French guide with zero accentsLanguage orthography
Orphaned guidesMath AI SL created but not wired to homepageNavigation entry points
AP nav hiddenAP French live but nav bar still IB-onlyProgram-level navigation
Commit formatCloses t-XXXX instead of Closes: [t-XXXX]Tool-specific syntax
Placeholders shipped”Coming soon” sections sent to productionContent completeness
Homepage hardcoded4 subjects hardcoded, new subjects invisibleDynamic vs static rendering

Fix Pattern

Always the same four steps:

  1. Research — agent must actively investigate the new domain before acting
  2. Configure — document domain-specific rules in config (workflow-config.json, CLAUDE.md)
  3. Validate — automated checks catch violations (content-validate.mjs, ux-review.mjs)
  4. Regress-test — add to the test suite so it can’t recur

Design Principle

Never trust an agent to know what it doesn’t know. The pipeline must include:

  • A research step before creation (creator reads CLAUDE.md, researches target language)
  • A validation step after creation (validator checks against known rules)
  • A rendered verification step (UX reviewer checks the actual output)

Each layer catches different classes of blindness. No single layer is sufficient.

Relation to Multi-Agent Pipeline

The 6-agent pipeline in apibstudyguide was designed to catch errors through layered review (creator → validator → reviewer → ux-reviewer). Agent domain blindness is the class of errors that slip through ALL layers because none of them have the domain knowledge to detect the issue. The fix is adding domain-specific rules that each layer can check against.

See also: multi-agent-architecture, source-agent-foreign-language, synthesis-project-patterns