Context Engineering 2 - как собрать AI-агента, который не тупит
In this video, we'll discuss why you can't work with an agent the same way you would with a regular chat. When an AI simply responds, a good request is often enough. Once an AI receives a role, tools, sources, memory, and the right to take steps, it needs a proper task definition. We'll explore how a model differs from a chat, assistant, and agent, why an Agent Capability Card is needed, how to create an agent role card, and how context determines the role, sources, constraints, status, and result verification for the model. I'll also show where the agent should stop: if there's insufficient access, sources are arguing, the task concerns money, personal data, legal promises, or sending a message out. After watching the video, it will be easier to assign a task to an AI project manager, a ticket review agent, or any working agent: with sources, boundaries, approval, and a verifiable result. 🎓 AI OS - materials + more useful information here: https://t.me/tribute/app?startapp=sJyg 🔗 Telegram channel: https://t.me/Sprut_AI 💻 GitHub: https://github.com/AlekseiUL Timecodes 00:00 Why one prompt is no longer enough 00:59 Prompt as a mini-regulation 01:46 Model, chat, assistant, and agent 02:42 Why an agent is dumb without context 03:30 Chat response and agent workflow 04:51 Agent Capability Card: what an agent can actually do 06:14 What blocks does a controlled agent consist of 07:21 Bad and normal problem statements 08:48 Why do we communicate with an agent, not a model 09:07 Agent Role: Job Title, Not Image 10:25 Role Card: How to Describe a Future Agent 11:16 Tools, Memory, and Approval 12:13 What an Agent Should Leave Behind 12:48 Stop Rules for an Agent 13:03 What an Agent Sees Before Starting a Task 13:49 Context Engineering as an Agent's Desktop 14:51 10 Layers of Work Context 15:38 Why Giving an Agent Too Much Information is Harmful 16:30 Work Context: Selection, Order, and Verification 17:21 Source of Truth and the Conflict Rule 18:06 Where an Agent Should Stop 18:49 Context Inventory: Context Map 19:34 An Example of an Initial Ticket Review Agent 20:22 A Practical Agent Assembly Flow 21:13 Summary: An Agent Needs a Working System #AI #AI #AIagents #AIAgents #PromptEngineering #ContextEngineering #LLM #ChatGPT #Automation #AIOperations
Why Prompt Engineering Is No Longer Enough: Context Engineering Begins
Context Engineering 3 - источники истины, RAG и память агентов.
Новое интервью Карпатого: мы создаём не разум, а призраков без контроля
ChatGPT-5.6 ПОБЕДИЛ Claude? Честный разбор на БИЗНЕС задачах
How RAG and LLM Wiki Work. Explained in 8 Minutes.
7 Authentication Concepts Every Developer Should Know
Spec-Driven Development: How to Write Production Code with AI Agents | Sergey Belov, Yandex
Stop Prompting Claude. Use Karpathy's Method Instead.
Claude + Obsidian: How to Build an AI System with Perfect Memory
CLAUDE IN 1 HOUR | How to Use Claude Better Than 94% of People | The 2026 Claude Guide
System Design Interviews: Mistakes That Sink You
How to Understand RAG in 18 Minutes, Even if You've Never Heard of Embeddings
Git-based skills: the new memory for AI agents. My experience
Не будь оператором LLM – освой Loop Engineering с агентами
7 скиллов для Codex и Claude Code на каждый день
Карпатый Wiki Вместо RAG — Полный Obsidian Сетап Для Новичка
Context Engineering, Part 4: Context Window, Evals, and AI Agent Safety
You Can Learn AI Agent Memory System In 12 Min | Semantic & Episodic Memory, RAG, Vector Database
Универсальная ИИ Система Для Всех (Wiki LLM, Obsidian, Supabase)
