Context Engineering for Long-Running AI Agents, Explained in 24 Minutes

Models now advertise context windows of a million tokens, but the amount of context they use well is much smaller. In this class, Dr. Aki Wijesundara (TAI Labs) explains long context engineering: why accuracy drops as the window fills up, what the research says about it, and how to manage context for AI agents that run for hours. What you will learn: The difference between prompt engineering and context engineering What competes for space in the context window, and why agents fill it so fast Why long contexts lose accuracy: the attention budget, lost in the middle and context rot Why the advertised context length is not the usable length Four ways long contexts fail: poisoning, distraction, confusion and clash Four moves for managing context: write, select, compress, isolate How to write system prompts, tools and examples that do not waste tokens Compaction, structured notes and subagents for long-running agents How to design around the KV cache so repeated context is cheaper and faster A seven-point checklist for your own agents Links mentioned in this video Lost in the Middle (Liu et al.): https://arxiv.org/abs/2307.03172 Context Rot (Chroma Research): https://research.trychroma.com/contex... NoLiMa: long context evaluation: https://arxiv.org/abs/2502.05167 Context Engineering for Agents (LangChain): https://blog.langchain.com/context-en... Effective context engineering for AI agents (Anthropic): https://www.anthropic.com/engineering... Context Engineering for AI Agents: Lessons from Building Manus: https://manus.im/blog/Context-Enginee... Less Context, Better Agents (Lodha et al.): https://arxiv.org/abs/2606.10209 How we built our multi-agent research system (Anthropic): https://www.anthropic.com/engineering... Don't Build Multi-Agents (Cognition): https://cognition.ai/blog/dont-build-... Chapters 0:00 What this lesson covers 0:33 What long context engineering is 1:23 From prompts to context 2:08 What competes for the context window 2:52 Why more tokens is not more context 4:13 Lost in the middle 4:48 Context rot, measured 6:29 Advertised length is not usable length 8:13 The 2026 picture for frontier models 9:26 Four ways long contexts fail 10:21 Four moves: write, select, compress, isolate 11:28 System prompts, tools and examples 13:18 Load context just in time 14:11 Compaction, structured notes and subagents 15:45 Design around the KV cache 18:33 Steering attention in a long task 19:27 Less context, better agents 20:52 One thread or many agents? 21:30 A working checklist About TAI Labs We help teams and professionals become AI-native. 10,000+ professionals trained. Official Claude Partner. Learn live with us on Maven (free Lightning Lessons and bootcamps): https://maven.com/tailabs Website: https://tailabs.ai #ContextEngineering #AIAgents #LLM #AIEngineering #ClaudeCode #PromptEngineering