Agent Context Workshop

A context layer for coding agents — built, then measured.

Most coding agents explore a codebase by grepping it. This workshop builds the alternative: a code knowledge graph an agent can traverse instead — then runs an honest eval to see whether it actually helps, and whether the agent uses it.

The worked example is qdrant (a real Rust vector database). We extract its structure into a Neo4j graph and give an agent graph tools (search, blast_radius, implementors, …). Then we pit a graph agent against a grep agent on 30 gold-verified questions.

The honest result: on single-shot accuracy they roughly tie (graph 21/30, grep 19/30). The graph wins on what matters in production — determinism (right on every run: graph 19/30 vs grep 11/30, and 12/15 vs 6/15 on relationship questions — grep re-derives the connections each run and flips), efficiency (~3× fewer input tokens), and the metric nobody reports: adoption (does the agent reach for the graph at all — 34% → 72% with the right nudge).

The two mental models

Three rungs of building the context layer (each notebook walks them):

Three levels of context engineering (what the eval compares): L1 grep → L2 index + search → L3 graph traversal.

Quickstart

Prereqs: uv, Docker (optional — only for notebook 01's graph), and an OpenRouter key (optional — only for live LLM calls; the eval is cached).

uv sync                       # install (slim — one provider, via OpenRouter)
scripts/setup_data.sh         # hydrate the committed qdrant seed (~2MB) into data/
cp .env.example .env          # add OPENROUTER_API_KEY for live calls (optional)
uv run jupyter lab

Then — new here? jump to notebook 02 (the eval; no Docker, no key needed). The full set, in order:

• notebooks/00_start_here.ipynb — a 30-second preflight: confirms your install + data are ready and routes you. Run it first.
• notebooks/01_posting_list_context.ipynb — bu