What is agent-context-workshop?

Agent Context Workshop is a hands on workshop, built for the AI Engineer World's Fair 2026, that teaches how to give coding AI agents a better way to understand a codebase than simply searching through it with grep. The alternative it builds is a code knowledge graph, a structured map of a codebase's functions, types, and relationships that an agent can traverse directly. The worked example uses qdrant, a real Rust based vector database, extracting its structure into a Neo4j graph and giving an agent tools to search it, trace the effects of a change, or find implementations of an interface. The project then runs an honest comparison between an agent using this graph and a plain agent using grep, tested against 30 verified questions. On a single attempt the two approaches score about the same, 21 correct out of 30 for the graph versus 19 for grep. But the graph pulls ahead on things that matter more in practice: it gets the same question right every time it is asked, while grep sometimes flips its answer because it has to work out the connections fresh each run. The graph also uses roughly three times fewer input tokens, and with the right prompting the agent actually chooses to use the graph far more often, rising from about a third of the time to nearly three quarters. The project is organized as a series of Jupyter notebooks that build up the idea in stages, from basic code parsing, to resolving types and relationships, to adding meaning with the help of a language model. Newcomers are pointed straight to the notebook that runs the comparison, which works without Docker or an API key since its results are cached. Setup uses the uv Python tool to install dependencies, a script to load a small prepared copy of the qdrant codebase, and an optional OpenRouter API key if you want to run live model calls instead of the cached results. The project is released under the Apache 2.0 license, and the example codebase it studies, qdrant, is also Apache 2.0 licensed.