Glossary

Key terms from the book in one line; details and context – in the note at the link.

Agent Architecture

  • ReAct – reasoning → action → observation loop, the main operating mode of the agent. [ReAct Cycle]
  • Harness – infrastructure around the model: context, tools, and three protective layers (restriction, verification, correction). [Harness Engineering]
  • MCP – Model Context Protocol, an open standard for connecting tools and data to models. [MCP Protocol]
  • Sidecar – lightweight parallel LLM call, gatewaying each tool call with structured data. [Sidecar Mechanism]
  • A2A – Agent2Agent (Google, 2025): agent-to-agent interaction protocol via tasks and artifacts. [A2A Protocol]
  • Computer Use – agent controls the graphical interface like a human: screenshot → reasoning → action. Computer Use
  • VLA – vision-language-action: model for robots, slow planning plus fast control. [VLA Models]
  • KV Cache – cache of computed key-value pairs of the attention mechanism: past tokens are not recalculated at each step. KV Cache
  • Prompt Cache – inter-request cache of the same prompt prefix at the API level. Prompt Cache
  • Chat Template – transforms API messages into a linear stream of tokens with service tags. Chat Template
  • RAG – Retrieval-Augmented Generation: retriever finds knowledge base fragments, generator answers based on them. RAG
  • Embedding – mapping text to a vector space, where semantically close is nearby. [Dense Embedding]
  • BM25 – sparse search for exact word matches: term frequency saturation and document length normalization. [Sparse BM25 Search]
  • Hybrid Search – dense and sparse engines in parallel, rank merging and neural reranking. [Hybrid Search]
  • GraphRAG – knowledge as a graph of entities and relationships: multi-hop inference and disambiguation. GraphRAG

Training and Evaluation

  • LLM-as-a-Judge – model evaluates quality according to an expert rubric; calibration on a gold set is required. LLM-as-a-Judge
  • RLHF – Reinforcement Learning from Human Feedback: SFT → reward model → PPO with KL penalty. RLHF
  • GRPO – RL without a value network: trajectory advantage – its position within a sampled group. GRPO
  • RLVR – RL with verifiable rewards: signal is issued by a rule-verifier, not a trained model. [RLVR and Verifiable Rewards]

Related: [Agent = LLM + Context + Tools], [Agent Formula], [Context Engineering]

// связанные заметки
←About the Project →Computer Use →KV Cache →Prompt Cache →Chat Template