SACE — Small Agent, Context Engineered

Can a small model become useful if we stop one-shotting it and instead let it walk a pre-structured tree of knowledge — one small, local decision at a time?

SACE is a research playground for that question. The bet: spread the data, spread the decisions. Don't hand a tiny model a flat ocean of chunks and ask it to find the answer. Hand it a dendrogram — a hierarchy where the topic narrows level by level — and let it zoom in, the way you zoom into Google Maps: continent → country → city → street.

The idea in one paragraph

Knowledge has hierarchy and boundaries, and those boundaries are easy to spot from a top-down view. Instead of asking the model to construct structure out of flat text, we give it a structure that already exists and ask it to use it. At each step the model sees only the current node and its direct children, makes one local routing decision (descend(child_id) or stop), and moves on. Context accumulates slowly, like an illumination process — general first, then specific. A dendrogram also makes a second point visible: a large system is built from many small ones. Even if the small model has never heard of a term, it can recognize what that term is made of, because the ingredients are its children.

The full thesis lives in the docs:

• docs/00-overview/05-the-idea.md — why spread > one-shot
• docs/00-overview/06-progressive-zoom.md — the Google-Maps zoom metaphor
• docs/00-overview/07-emergence-from-ingredients.md — small systems compose into larger ones

The debugger

Because the reasoning is the product, the UI gives equal real estate to the trace. Two synchronized views — chat-style (turn-by-turn list) and tree-overlay (route on the map) — show the same run from different angles. Pick any past message and replay how the agent got there.

More on this in docs/00-overview/01-vision.md and docs/06-frontend/04-debug-panel.md.

Quick start

# Backend
uv sync --group dev
make dev-backend   # http://localhost:8000

# Frontend (separate terminal)
cd frontend && npm install && npm run dev   # http://localhost:5173

Layout

• backend/sace/ — Python package (FastAPI, LangGraph, schemas)
• frontend/ — React + Vite SPA
• data/trees/ — example knowledge trees (JSON, seeded into SQLite at boot)
• docs/ — design, architecture, and the idea writeups
• assets/ — figures used in the docs

Make targets

Target	Description
make dev-backend	Uvicorn with reload
make dev-frontend	Vite dev server
make test	pytest
make seed	Load trees from data/trees/
make types	Regenerate frontend types stub

Status

Experimental. Mini-tier models only (gpt-4.1-mini, gpt-4o-mini) — the constraint is the experiment. If small models can't route a well-shaped tree, that itself is the finding.