Advanced Context Engineering: Systems, Evaluation & Research

Advanced Context Engineering: Systems, Evaluation & Research

Build, measure and test context systems for text, knowledge and agents, and check ideas from research yourself.

What you’ll learn
Build context you can verify: extract text from PDFs, split documents, and assemble task packets whose claims trace to a file and a line.
Compare word search (BM25), embeddings, rank fusion and reranking on the same questions, and choose a method from measured evidence.
Chain dependent searches, add source-grounded context to chunks, and audit a small NetworkX graph against current permissions.
Manage long context with budgeted selection, checked shortening, handoffs after compaction, and caches whose keys include content and permissions.
Build memory with provenance, validity dates and lifecycle rules in SQLite, and test it across separate conversations.
Build a local MCP server with typed inputs, pagination and failure handling that does not invent answers.
Run long tasks with explicit state and checkpoints, resume after a stop, and measure whether several agents beat one.
Detect instructions hidden in data, keep sources through handoffs, and enforce permissions in code, testing for leaks and for over-blocking.
Evaluate systematically: test checkers against hand labels, catch regressions, run ablations, and improve instructions from feedback.
Read a research paper for its claim and limits, test its idea against fair baselines, and finish a system on a new domain with a sealed test.

Requirements
Context Engineering, Part I, or equivalent comfort with the Codex CLI: working in a terminal with folders and files, and checking what a tool actually did. Lesson 1 assumes Codex is already installed and signed in.
A Codex CLI account of your own. The lessons were recorded with GPT-6-Luna at medium reasoning, and results on a different setup may differ. Model usage counts against your account, and the course does not promise a fixed cost.
A Linux (or other POSIX) shell, Python 3.12 and the uv package manager, which lesson 2 walks through installing. Git is added in lesson 6, and free Python packages such as pytest are installed as lessons need them. Only the Linux route was demonstrated.
Willingness to read and run Python line by line. No programming knowledge beyond Part I is assumed, and Python starts in lesson 2, but the pace is fast and you may need to stop and look things up.
Optional: an Alibaba Model Studio account with a Singapore-region API key for the live embedding and reranking calls in lessons 15, 16 and 18. Saved responses are supplied so you can follow without an account; live calls may be billed.
Time for real work: lessons run from about 10 to about 42 minutes, and the first half of the course is the heaviest.

Description
This course contains the use of artificial intelligence.

Advanced Context Engineering is the advanced course that follows Context Engineering, Part I. It follows one idea through 60 lessons in 12 modules, about 19 hours in total: the context an AI agent works with is data that you assemble, check and measure. You start from a small baseline and finish with a complete system that you test on a sealed test set and defend with evidence.

Along the way you prepare documents and PDFs you can trust, assemble task packets whose claims trace to a file and a line, and compare word search, embeddings, rank fusion and reranking on the same questions. You build chained searches and a small graph, manage long context with budgets, handoffs and caches, and build memory with sources, dates and SQLite. You then build a local MCP server, run long tasks that can stop and resume, measure whether several agents help, plant instructions in data, enforce permissions in code, and build evaluation suites that catch regressions and show which component actually helped. The last modules cover reading a research paper and testing its idea, and a final project: a staff assistant for a fictional warehouse handbook.

Every lesson follows the same pattern: a short refresh, slides that explain the new idea, a continuous demonstration in Codex CLI, a close reading of the results, and a recap. Almost every lesson ends with an exercise and a solution to compare against. Lessons come with starter files and saved results and, in the later lessons, offline checks and scripted stand-ins that make no model calls, so you can inspect what was recorded before spending anything. The demonstrations use Codex CLI 0.156.1 with GPT-6-Luna at medium reasoning, and all data is synthetic, drawn from fictional organizations.

It is written for graduates of Part I and for people who write code or are ready to learn as they go. No programming knowledge beyond Part I is assumed, and Python is introduced from lesson 2, but the pace is fast and the code is read line by line. If Context Engineering or the Codex CLI is new to you, start with Part I.

What it is not: it does not promise research training equivalent to a doctorate. GraphRAG, GEPA, ACE and Recursive Language Models are taught as principles and as small, clearly labeled adapted demonstrations, not as reproductions of the papers. Most results come from small synthetic experiments with one recorded run, so they show a method and not a benchmark, and the safety lessons demonstrate mechanisms without promising protection in production. OCR of scanned documents, real-time audio and video, and model training are outside the course. Model usage counts against your own account, and the course does not promise a fixed cost.

Who this course is for
Graduates of Context Engineering, Part I who want to go from prompts and folders to systems they can measure and improve.
Developers and engineers who build assistants, retrieval or agent systems and want habits for evaluation, permissions and testing.
Learners without a programming background who are ready to learn Python as they go and to stop and look things up.
People who want to read a research paper and test one of its ideas themselves in a small, fair experiment.

AVCEDENTECOATRECETRSYATREREDVELRUTE

you must be registered member to see linkes Register Now