Context Engineering
by DeepLearning.AI and Chroma
Build agent harnesses in plain Python and measure how search, pruning, compaction and memory change what the model sees.
Overview
Context Engineering is a DeepLearning.AI course made with Chroma, announced on October 20, 2026, with the course page showing an update on October 22, 2026. Jeff Huber, co-founder and CEO of Chroma, teaches it, and Drew Breunig (CEO of CMPND and editor of an upcoming O'Reilly handbook on context engineering) appears in a four-part guest segment. The page lists 18h 56m of material: 37 video lessons, 8 code examples and 6 graded assignments for Pro members. It has four modules. Module 1 defines context and its components, covers context failure modes and context rot (the effect Chroma measured across 18 models in its 2025 research report), and takes apart the components of an agent harness through labs on a customer support agent. Module 2 covers agentic search: lexical, semantic, hybrid and structured search strategies, tool registries that let an agent search for the right tool, data preparation for retrieval, and how to evaluate search. Module 3 covers context management, using a research agent to work through pruning, compaction and scratchpads, with labs that inspect the context window directly. Module 4 covers agentic memory: reading and writing memories for an email agent and evaluating whether they help. The labs build harnesses in Python without an agent framework, so the techniques carry over to whatever stack you use. Videos are free to watch; graded assignments and the certificate need DeepLearning.AI Pro.
At a Glance
- Topic
- Agentic
- Level
- Intermediate
- Format
- Course
- Cost
- Freemium
- Duration
- ~19 hours, self-paced (37 video lessons, 8 code examples, 4 modules)
- Provider
- DeepLearning.AI and Chroma
- Hands-on
- Yes — code/exercises
- Certificate
- Available
What You’ll Learn
- ✓Diagnose context failure modes and context rot inside a working customer support agent
- ✓Identify the components of an agent harness and how each one shapes the model's context
- ✓Implement lexical, semantic, hybrid and structured search strategies as agent tools
- ✓Build a tool registry so an agent can search for the right tool instead of loading all of them
- ✓Evaluate agentic search with precision and recall and improve it through chunking, code parsing and metadata
- ✓Apply pruning, compaction and scratchpads to a research agent and compare runs with and without each
- ✓Read and write persistent memories for facts, rules and artifacts, and decide when to add, update or skip one
- ✓Evaluate whether an agent's memory actually improves its results using traces and evals
Highlights
- •Ten one-hour labs build agent harnesses in plain Python with no framework, so the patterns transfer to LangGraph, the Claude Agent SDK or a custom loop
- •Taught by the Chroma CEO whose team published the 2025 Context Rot report on how performance degrades as input length grows
- •Every module ends with an evaluation lesson, so you measure each context technique instead of assuming it helps
- •Includes a four-part guest series with Drew Breunig, who edits an O'Reilly handbook on context engineering
- •Newly released in October 2026 and covers tool registries and agent memory, which most older RAG courses skip
Who It’s For
Best For
- ✓Engineers whose agents degrade on long tasks or with large tool sets
- ✓Developers moving from basic RAG pipelines to agentic search
- ✓Teams designing memory for multi-session assistants
- ✓Builders who want framework-independent harness skills
Prerequisites
- •Intermediate Python, including functions and classes
- •Familiarity with LLM applications, tool calling and basic retrieval (DeepLearning.AI suggests its Agentic AI and RAG courses as refreshers)
FAQ
What is Context Engineering?
Context Engineering is an intermediate DeepLearning.AI course, built with Chroma and taught by Chroma CEO Jeff Huber, for developers who already build LLM apps with tool calling and basic retrieval. You will be able to build agentic search, context pruning, compaction, scratchpads and persistent memory from scratch in Python, and evaluate each one with traces and evals.
Is Context Engineering free?
Context Engineering offers free content, with paid options for certificates or premium features.
What level is Context Engineering for?
Context Engineering is aimed at a intermediate audience. Recommended background: Intermediate Python, including functions and classes, Familiarity with LLM applications, tool calling and basic retrieval (DeepLearning.AI suggests its Agentic AI and RAG courses as refreshers).
How long does Context Engineering take?
Expect roughly ~19 hours, self-paced (37 video lessons, 8 code examples, 4 modules). Most learners work through it at their own pace.
What will I learn from Context Engineering?
You'll learn: Diagnose context failure modes and context rot inside a working customer support agent; Identify the components of an agent harness and how each one shapes the model's context; Implement lexical, semantic, hybrid and structured search strategies as agent tools; Build a tool registry so an agent can search for the right tool instead of loading all of them; Evaluate agentic search with precision and recall and improve it through chunking, code parsing and metadata; Apply pruning, compaction and scratchpads to a research agent and compare runs with and without each; Read and write persistent memories for facts, rules and artifacts, and decide when to add, update or skip one; Evaluate whether an agent's memory actually improves its results using traces and evals.
Topics
Sources
This page was written from 4 sources, 3 on domains other than deeplearning.ai.