DeepLearning.AI Launches Context Engineering Course, Teaching Agent Context Management

DeepLearning AI has launched a new course, "Context Engineering."

The course treats "context engineering" as the main thread for keeping Agents reliable and token-efficient on complex multi-step tasks: it's not about relying on a larger window or rewriting a single prompt, but about managing what the model sees, how it's arranged, and when to retrieve and remember. The course builds a harness from scratch in Python and uses traces and evals to quantify every change! https://t.co/UKFzLhFAgb

Two instructors @trychroma CEO @jeffreyhuber - Chroma has been acquired by Replit CMPND CEO @dbreunig - author of the Context Engineering Handbook

Why learn this

The longer an Agent runs, the more easily its context fills up with repetitive, contradictory, and outdated information. Models perform worse on long contexts, and Jeff Huber and the Chroma team call this phenomenon Context Rot. A larger window can't fix it, and simply changing the prompt can't fix it either; what needs fixing is the context itself: putting truly relevant information in the right format at the right time.

Three capabilities you'll practice · Agentic Search: multi-tool, multi-source data retrieval; use traces to improve tool selection, and close the loop with evals. · Context Management: use pruning, compaction, and scratchpads to organize the information a task needs. · Agentic Memory: store user feedback and the like as reusable memories, retrieve them when a task needs them, and add or modify them across sessions as things change.

Who it's for

People building AI Agents who want complex multi-step tasks to be more stable and more token-efficient. Prerequisites: intermediate Python (functions and classes), familiarity with LLM applications, tool calling, and basic retrieval. The page suggests warming up with @DeepLearningAI's Agentic AI and RAG courses first.

# Course structure (four modules)

Module 1 | Introduction to Context Engineering Opens with a conversation between Andrew Ng and Jeff Huber. It explains what context is made of, common failure modes, what Context Rot is, and how to use the harness's hooks and traces to detect and prevent problems. Labs include: context failure modes, dissecting a customer service Agent, and hands-on shaping of its context. Drew Breunig's first guest segment is about 7 minutes.

Module 2 | Agentic Search Uses a coding Agent as the example. Covers lexical, semantic, hybrid, and structured retrieval; writing good tool descriptions and using a tool registry to load tools on demand; improving retrieval with chunking, code parsing, and metadata, and measuring it with precision / recall. Includes labs such as building a search system and a tool registry; Drew's second guest segment is about 5 minutes.

Module 3 | Context Management Uses a research Agent as the example. First get a clear view of what's in the context window, then practice pruning (hooks for deduplication, extraction, truncation, etc.), compaction (compressing long histories into structured summaries), and scratchpads (moving full text and task state outside the window). Compare run differences before and after turning each operation on / off, and do context management evals. Drew's third guest segment is about 4 minutes.

Module 4 | Agentic Memory Uses an email Agent as the example. Distinguish memory records that can persist across sessions, such as facts, rules, and artifacts; decide whether each new piece of information should be added, updated, or skipped; and evaluate whether memory actually saves users from repeating themselves. There are labs for reading and writing memory; Drew's fourth guest segment is longer (about 17 minutes), followed by the course wrap-up.

Each module has graded assignments and optional materials; the labs are mostly code exercises on the order of about 1 hour (as noted on the page).