Chroma | Context Engineering
Conversations with practitioners and researchers building with large language models, cutting through the hype to examine what actually works in production.
Context engineering, agent architecture, model selection, and the gap between benchmark scores and real-world performance. Grounded technical perspectives for builders navigating the practical challenges of AI engineering.
Chroma | Context Engineering
Drew Breunig
Jeff Huber sits down with Drew Breunig to talk about the origins of Context Engineering and more.
Drew has a wide range of his writing about AI on his website: https://www.dbreunig.com/
00:36 Why write about AI? (Writing as a searchable index)
01:28 The two buckets of AI writing: Hype vs. Research
04:08 The Gemini 1.5 Paper & Pokemon: The birth of Context Engineering 06:50 The "Karpathy Effect" on Context Engineering
08:17 Benchmarks, Model Cards, and the "Agent Harness"
11:41 The Weightlifting Metaphor for AI Benchmarks
14:02 Testing Opus 4.5: Building internal tools in one shot
15:54 Models are untapped: The gains are in the harness
17:05 Why isn't there a standard Context Engineering harness?
19:20 The "Hello World" Experiment: Testing Agent Frameworks (LangChain, Crew, etc.)
21:42 Compact and Grep vs. File Systems
23:45 "Naked" tool calls vs. Frameworks
24:50 The GPT-8 Thought Experiment: Why Software Engineering still matters
27:12 Compound AI Systems: What agents can learn from Data Pipelines
31:00 Reliability is the bottleneck (The MAP Report)
36:00 Token Speed: When code generates faster than humans can read (Groq/Cerebras)
41:00 The UX of Multi-Agent Systems (The "Starcraft" problem)
43:57 ChatGPT Deep Research: The Shopping Use Case
46:25 Building Trust: Agent Design as Client Services
49:30 Continual Learning: Weights vs. Context/Memory
51:50 The problem with "Black Box" memory (The Chocolate Example)
56:30 The need for "Modes" (Work vs. Home context)
--
Chroma is the open-source AI application database. Batteries included.
Embeddings, vector search, document storage, full-text search, metadata filtering, and multi-modal. All in one place. Retrieval that just works. As it should be.
Try it today:
https://trychroma.com/cloud