◆ Supervised Worker
Ingests a policy update or system rollout and authors tailored learning modules, quizzes and job aids per role, refreshing them when the source material changes. A subject-matter judge agent gates every module before publish.
Memory
Working The source material and module being authored.
Semantic Curriculum map, learning objectives, accessibility standards.
Procedural Module templates and assessment patterns refined by SME-judge feedback.
Store Content library + curriculum graph
Orchestration
pipeline MCPA2A
Harness · Managed Agents … drafting session with a subject-matter judge agent gating publish.
Tools
⌕ Policy + system doc retrieval Retrieval { } LMS (learning management system) API ⇄ SME review judge A2A ›_ Authoring sandbox Code exec
Evals & guardrails
- Subject-matter judge agent approves every module before publish … accuracy is non-negotiable.
- Agent-as-judge check that assessment items map to stated learning objectives.
- Accessibility (WCAG) validation on generated content.
Offline reflection
Replays learner assessment results to find weak modules and self-revises the explanations that consistently confuse people.
Frontier edge
- ▲Multimodal authoring: reads the source policy PDFs, system screenshots and demo screen-recordings natively, then generates interactive walkthroughs from them.
- ▲Eval-gated continual learning: assessment-failure patterns drive self-edits to the weakest explanations between cohorts, each revision regression-checked by the SME judge before publish.
- ▲Self-improving fleet: a clearer explanation that lands in one role's module is shared across the curriculum graph so sibling courses inherit it.
In numbers
under 40 min
Median module authoring time
median under 24h after a policy change
Training-freshness lag
Handoffs
Across ⇢ Compliance for mandatory-training content sign-off