کتاب راهنمای مصور عامل‌های هوش مصنوعی

مقدمه

What Is an AI Agent? • Large Language Models • Reasoning Large Language Models • Augmenting the Large Language Model • An Agentic System • Specializations • Multi-Agent Collaboration • The Multi-Modal Agent • The Coding Agent • The TinyAgent • Summary

مدل‌های زبان بزرگ

Part 1: What You Should Know About Large Language Models • Input and Output Tokens • From Language Modeling to Powering Agents • The TinyAgent • Training a Large Language Model • The Transformer Architecture • Part 2: A Deeper Dive Into Large Language Models • How Self-Attention Works • KV-Caching Revisited • More Efficient Self-Attention • Mixture of Experts • Summary

مدل‌های زبان بزرگ استدلالی

The Paradigm Shift from Train-Time Compute to Test-Time Compute • Train-Time Compute Versus Test-Time Compute • Scaling Laws • Categories of Test-Time Compute • Reward Models • Prompt Engineering • Search Against Verifiers • Self-Consistency • Best-of-N Samples • Modifying Proposal Distribution • Supervised Fine-Tuning • Reinforcement Learning • Native Reasoning • Emerging Directions in Reasoning Research • Reasoning in Multi-Modal LLMs • Efficient Reasoning • Reasoning in Latent Space • Summary

حافظه

Types of Memory • Short-Term Memory • Conversation Memory • Trimming • Summarization • Long-Term Memory • Retrieval-Augmented Generation • Agentic Retrieval-Augmented Generation • Context Engineering • Context Engineering for Multi-Agent Systems • Optimizing the Context • Context As the Specification • Summary

استفاده از ابزار، یادگیری و پروتکل‌ها

Tool Usage • Tool Creation • Tool Definition • Tool Selection • Tool Calling • Tool Output Processing • TinyAgent with Tools • Tool Learning • In-Context Learning • Supervised Fine-tuning • Reinforcement Learning • TinyAgent with Native Tool-Calling Capabilities • Model Context Protocol • Core Components • The MCP Flow • Skills • The SKILL.md • The Bundled Resources • Summary

برنامه‌ریزی و بازتاب

Planning • Task Decomposition • Action Sequencing • Agents That Continuously Improve • Reflection • Self-Improvement • Summary

ارزیابی عامل‌ها

Public Benchmarks and Leaderboards • Coding and Software Engineering • Tool Use • Software and Computer-Use Environments • Real-World Task Completion • Reasoning and Knowledge • Other Groups of Benchmarks • Reading Benchmark Scores Critically • Outcome Evaluation: Did the Agent Get the Right Output? • Human Evaluation • Automated Evaluation • Trajectory Evaluation: Did It Get There the Right Way? • Reliability: Does It Succeed Every Time? • Measuring Capability with pass@k • Measuring Reliability with pass^k • Safety: Does It Avoid Harm? • Building Your Own Evals • Summary

سیستم‌های چندعاملی

The Multi-Agent System • Orchestrating Agents • Patterns • General-Purpose Frameworks for Collaborative Task-Solving Patterns • Communication Protocols • Agent Society • Emotional Intelligence • Simulations • Interactive Simulacra of Human Behavior • Deep Research Agents • Toward an AI Co-Scientist • Agent Laboratory • Summary

درک چندوجهی

What Are Multi-Modal LLMs? • Encoding Modalities • Text • Images • Audio • Video • Many-Modality • Connecting Modalities • Projection-Based Connector • Query-Based Connector • Fusion-Based Connector • Advantages and Disadvantages • TinyAgent • Summary

عامل‌های کد و مدل‌های زبان بزرگ کد

Users and Builders of Code Agents and Large Language Models • Building Code Agents • Code Agents to Serve Non-Coders • Code Tools • Context Management • Code Agents for Software Engineering • Task-Specific Workflows • Building Code LLMs • Summary