design:os · pedagogy · Agent Teacher
A full school year of Grade 1 mathematics, authored with an evidence-graded pedagogy library.
Agent Teacher is a pedagogy knowledge library graded by evidence, with a course-design workflow and a publication pipeline that validates before it applies. Its first worked artefact: a Vietnamese Grade 1 mathematics course, 35 weeks and 105 period sheets, checked by an executable audit. A draft for teachers to appraise, not an approved textbook.
Làm quen: mỗi vật đếm một lần
- Học cụ
- Khay và 3 nắp chai lớn.
- Bài mẫu của cô
- Cô chạm lần lượt 3 nắp chai, nói một, hai, ba; số cuối cho biết có 3 vật.
- Luyện tập
- Trẻ lấy đúng 2 que, sau đó lấy đúng 4 que. Đáp án: 2 que; 4 que.
- Nhiệm vụ cuối tiết
- Có bao nhiêu nắp chai trên khay? Đáp án/tiêu chí: 3
One of the 105 period sheets, as authored: teacher demonstration, practice, answers, an end-of-period task with its criterion. Every sheet has these fields.
lessons · Tiết 1What it is
- Course
- 105 period sheets35 weeks × 3 periods × 35 min = 3,675 min; 12 topics
- Audit
- 24 / 24 requirement codes111 declared answers recomputed of 125 structured tasks
- Teacher material
- 8 worked lesson planstwo 10-point sample exams, an ongoing-evidence sheet
- Library
- 92 records · S0–S7ten numbered areas, one versioned ontology, evidence tiers
- Tests
- 98 tests · 11 modulesregistry, publication, rehearsal, course audit
- Runtime
- Python 3.10+no model calls; two dependencies for registry and publication
Why this, compared with a skills pack, a content generator or a syllabus
What Agent Teacher does that the alternatives do not, or do differently. Every line below is stated in the README with its boundary.
A worked full-year course
Vietnamese Grade 1 mathematics: 24 requirement codes decomposed from the national programme, 12 topics, 105 period sheets each with demonstration, practice, answers and an end-of-period task, eight worked lesson plans, two sample exams.
An audit that can fail
The course auditor reads the bundle: calendar, references, encoded scope and structured answers. 20 unit tests include deliberately broken variants (missing standards, dangling references, prerequisite cycles, wrong answers, carrying). It does not judge meaning or children’s learning.
Evidence is graded, not asserted
Every record carries verification, claim status, effect metric, applicability and review status as separate fields; legacy grades confer no expert review.
Publication that can refuse
The registry checks metadata, canonical IDs, aliases and typed references; the publisher validates a draft, locks the filesystem, and recovers an interrupted transaction explicitly.
No model in the loop
The tools make no external model calls; an optional provider must be selected explicitly. Nothing here certifies teaching competence or an approved textbook.
Where the alternatives win, stated plainly: education-agent-skills drops 165 skills straight into Claude Code, Codex or Hermes; AI-Teaching-Agent generates labs, exams and grading workflows; Pyragogy ships a hosted syllabus. Still open here: source appraisal, named expert review and a real-user pilot. Implementation status.
Run it
Audit the course bundle and run its tests from the project root. The auditor checks the calendar, references, encoded scope and structured answers; it does not verify meaning, images or what children learn.
git clone https://github.com/jangtrinh/design-os-pedagogy.git && cd design-os-pedagogy
.venv/bin/python -B tools/course_audit.py courses/vi-vn-grade-1-math
.venv/bin/python -B -m unittest discover -s tests -p 'test_course_audit.py' -vDesign a new course with the assistant prompt, the design protocol and the blueprint template; prepare a module without writing anything; verify the whole library.
.venv/bin/python -B tools/pedagogy_pipeline.py "Fraction teaching routine" --stage S2 --archetype methods --dry-run
.venv/bin/python -B -m unittest discover -s tests -v
.venv/bin/python -B tools/verify_links.pyA local teacher rehearsal on three authored fraction cases also ships (python3 run_rehearsal.py --port 0); it is an early build and stays out of this page’s showcase. Its contract.
The library map
Start with the master map. The authoring model connects knowledge, experience, diagnosis, practice, feedback, transfer and an implementation protocol.
00-system
Ontology, schema, epistemic policy and authoring contract
10-foundations · 20-stages
Foundations and education-stage guides, S0 through S7
30-pedagogy · 40-disciplines
Methods, assessment and disciplinary practice
50-practice-library · 60-evidence
Cases, source notes, claims and estimates
70-capabilities · 80-professor-development
Teaching capabilities and professional learning
90-agent-runtime
Runtime specifications, prompts and evaluation proposals
rehearsal
The implemented local fraction rehearsal
tools · tests
Registry, publication, local link auditing and executable tests
FAQ
Does it call a model?
No. The rehearsal, registry and publisher make no external model calls. Free text is stored, not semantically graded.
Does it certify teaching competence?
No. Passing software tests or a clean reference graph does not establish teaching effectiveness; authored examples are not observed classroom outcomes.
What is implemented, and what is a specification?
The Vietnamese fraction rehearsal, the registry and the publication pipeline are implemented and tested. The broader tutor runtime in 90-agent-runtime remains a specification.
What remains before a real pilot?
Source appraisal, named expert review and a real-user pilot are pending; the fraction pilot document lists the human work.
Which languages?
The rehearsal is Vietnamese; the README and the master guide exist in English and Vietnamese.
What is the license?
Not yet declared in the repository. Until a LICENSE file lands, no reuse terms are stated.
DESIGN:OS Ecosystem
Modular toolchain for AI coding agents: from interface CLI to deterministic animation, 3D CAD and robotics.