Descriptions:
Jeff Huber, co-founder and CEO of Chroma, delivered a talk at AI Dev 26 in San Francisco laying out a comprehensive framework for agentic search — the challenge of giving AI agents reliable, contextually appropriate access to information at query time. Chroma, makers of the open-source vector database with over 27,000 GitHub stars and 140 million downloads, has expanded beyond local search to a distributed cloud offering and a 20-billion-parameter open-source model called Context-1, purpose-built for agentic retrieval tasks.
Huber reframes AI agents as context-and-reasoning machines, drawing on OpenAI data showing over 45% of ChatGPT queries are fundamentally search-centric. He introduces a 2×2 framework mapping query complexity (simple vs. complex) against corpus size (small vs. large), arguing that modern agentic systems must handle all four quadrants — a design requirement that classic RAG pipelines routinely fail to meet. He distinguishes between read-path retrieval (what should the agent reference at runtime?) and write-path memory (where should derived knowledge be stored?), positioning both as equally critical for agents that perform reliably on complex, multi-step tasks.
The talk also revisits Chroma’s “context rot” research — the degradation in reasoning quality when language models operate over excessively long, unfiltered context — work Huber says is now regularly cited by both Anthropic and OpenAI in new model releases. Chroma is also credited with popularizing the term “context engineering” as a discipline distinct from prompt engineering, reflecting a broader industry shift toward treating context curation as a first-class engineering problem.
📺 Source: DeepLearningAI · Published May 20, 2026
🏷️ Format: Keynote Launch
