AgentFabric overview — authoring package¶
This is the reproducible authoring package for the AgentFabric technical/marketing
overview: a presentation member (20-slide deck) and a narrative member (long-form
document) built from one reviewed source of claims.
The audience assumption is stated in the deck and must not drift: a highly technical reader who is new both to AgentFabric and to fleet-scale multi-agent / multi-model orchestration. Nothing in either member may assume prior product knowledge.
Narrative order, which is deliberate:
| Slides | What it argues |
|---|---|
| 1–5 | Marketing first: what the system does, why a chat window is not an operating model, the four properties it guarantees, and the end-to-end loop |
| 6–10 | Technology re-use: where AgentFabric leverages NVIDIA technology, where it leverages open source, the five mechanisms it adds anyway, and the containment boundary |
| 11–19 | How the machine actually runs: actors and their single authority each, task lifecycle, bus coordination, fleet shape, model routing and spend, evidence, measured surface area, and stated limits |
| 20 | Close and adoption path |
Diagrams and flows carry the argument. Bullet lists are used in one place on purpose — the NVIDIA and open-source project inventories on slides 7 and 8 — because a list of load-bearing dependencies is the point of those slides.
Read in this order¶
- current deliverables — what exists right now
- deck specification — slide-by-slide presentation authority
- narrative specification — long-form structure authority
- source notes — the factual ledger every claim traces to
- QA ledger — what was checked, and how to re-check it
- SKILL.md — the regeneration workflow
Model-facing inputs: deck-authoring prompt and image prompts.
Builders¶
| File | Role |
|---|---|
build_deck.py |
deterministic python-pptx deck; native shapes only, no raster diagrams |
build_narrative.py |
deterministic python-docx narrative member |
render_slides.py |
slide rasterizer, contact sheet, and text-frame overlap report |
publish_google_workspace.py |
optional publication; requires explicit authorization |
regenerate.sh / regenerate_python.sh |
regeneration entry points |
Every diagram in the deck is built from native PowerPoint shapes, so the deck stays
editable and legible after a Google Slides import. There is no assets/ directory and no
generated hero imagery: a mechanism diagram is preferred over decoration everywhere the
two compete.
Generated .pptx / .docx artifacts are build outputs. Regenerate them; do not treat a
stale copy as authority. Publication to any external location requires explicit
authorization from the owner of that location.