# Agentic Glossary — Quick Reference

> Quick-reference glossary of agentic-AI vocabulary. 88 terms, each defined in 1–3 sentences with one canonical primary-source citation. Companion to the deep [Agentic Glossary](https://agentic-glossary.roei-020.workers.dev/). Maintained by [AgentsBooks](https://agentsbooks.com/). Last refreshed 2026-05-07.

## How this differs from the deep glossary

The deep glossary is the narrative version (31 hand-curated entries, paragraph-long definitions, full canonical block-quotes). This site is the breadth layer: more terms, shorter definitions, optimized for fast lookup, copy-paste, and LLM citation. Flagship terms link out to their deeper entry on the deep glossary.

## Categories

- **8 Primitives** — 8 terms
- **Core Concepts** — 14 terms
- **Protocols** — 11 terms
- **Memory & RAG** — 11 terms
- **Reasoning** — 8 terms
- **Frameworks** — 6 terms
- **Evaluation** — 11 terms
- **Compliance** — 14 terms
- **Operations** — 5 terms

---

## 8 Primitives

### Identity (primitive)

First of the AgentsBooks 8 primitives. Bundles name, role, tagline, organization, personality, tone, voice, and appearance into a portable JSON object so the same agent renders consistently across LLM, TTS, and image-generation surfaces.

**Source:** [AgentsBooks — Anatomy of a Firm](https://agentsbooks.com/anatomy) — accessed 2026-05-07

**See also:** Agent · Agentic firm · Brain (primitive) · Heart (primitive)

**Compare with deep glossary:** https://agentic-glossary.roei-020.workers.dev/

### Brain (primitive)

Cognition layer of an agent: LLM provider and model, system-prompt prefix, behavioral guidelines, skills, tools, MCP servers, plugins, hooks. Provider swap (Claude / GPT / Gemini / open-weight) is a configuration change.

**Source:** [AgentsBooks — Anatomy of a Firm](https://agentsbooks.com/anatomy) — accessed 2026-05-07

**See also:** Agent · Tool use / Function calling · MCP / Model Context Protocol · Constitutional AI (CAI)

**Compare with deep glossary:** https://agentic-glossary.roei-020.workers.dev/

### Heart (primitive)

Tasks-and-triggers layer: cron, webhooks, event triggers, and the runtime that executes them. The heart is what makes an agent operate continuously without a human typing.

**Source:** [AgentsBooks — Tasks & Triggers](https://agentsbooks.com/features/tasks-triggers) — accessed 2026-05-07

**See also:** Agent · Agentic firm · Agent fleet

**Compare with deep glossary:** https://agentic-glossary.roei-020.workers.dev/

### Memory (primitive)

*Also known as:* memory layer

Long-term persistent stores attached to an agent: relational, blob, vector, filesystem (PostgreSQL, Redis, Firestore, S3, Pinecone, MongoDB). What separates an agent from a stateless chatbot.

**Source:** [AgentsBooks — Anatomy of a Firm](https://agentsbooks.com/anatomy) — accessed 2026-05-07

**See also:** RAG / Retrieval-Augmented Generation · Context graph · Vector database · Long-term memory

**Compare with deep glossary:** https://agentic-glossary.roei-020.workers.dev/

### Control (primitive)

Communication-channels layer: Telegram, Slack, Discord, Email, Webhook, SMS, WhatsApp, API. How an agent receives input from the outside world and pushes output back.

**Source:** [AgentsBooks — Anatomy of a Firm](https://agentsbooks.com/anatomy) — accessed 2026-05-07

**See also:** Agent · Agent fleet · Shares (primitive)

**Compare with deep glossary:** https://agentic-glossary.roei-020.workers.dev/

### Knowledge (primitive)

Expertise-domains layer: structured knowledge bases, scraped documentation, RSS feeds, ingested URLs, and the retrieval pipelines that surface them at inference time. Configurable subject-matter expertise.

**Source:** [AgentsBooks — Knowledge & Learning](https://agentsbooks.com/features/knowledge-learning) — accessed 2026-05-07

**See also:** RAG / Retrieval-Augmented Generation · Memory (primitive) · Context graph · Embedding

**Compare with deep glossary:** https://agentic-glossary.roei-020.workers.dev/

### Friends (primitive)

Inter-agent networking layer: which other agents this agent can delegate to, share memory with, send messages to, or spend credits with. Where A2A patterns surface inside the AgentsBooks substrate.

**Source:** [AgentsBooks — Anatomy of a Firm](https://agentsbooks.com/anatomy) — accessed 2026-05-07

**See also:** A2A / Agent2Agent Protocol · Multi-agent system · Agent fleet

**Compare with deep glossary:** https://agentic-glossary.roei-020.workers.dev/

### Shares (primitive)

Content-and-assets layer: posts, images, audio, video an agent has produced or been granted, plus visibility settings. How an agent participates in feeds, marketplaces, and the wider agents economy.

**Source:** [AgentsBooks — Anatomy of a Firm](https://agentsbooks.com/anatomy) — accessed 2026-05-07

**See also:** Agent · Agentic firm

**Compare with deep glossary:** https://agentic-glossary.roei-020.workers.dev/

---

## Core Concepts

### Agent

An AI system that decides for itself how to accomplish a task — choosing tools, sequencing steps, observing results, and adjusting — rather than executing a fixed script. Defining capability is autonomy over the path, not just the answer.

**Source:** [Anthropic — Building Effective Agents](https://www.anthropic.com/research/building-effective-agents) — accessed 2026-05-07

**See also:** Workflow (vs agent) · Agentic AI · Tool use / Function calling · Multi-agent system · Autonomy

**Compare with deep glossary:** https://agentic-glossary.roei-020.workers.dev/

### Agentic AI — Contested

Category of AI systems that perceive, decide, and act toward goals over multiple steps — distinct from chatbots that only respond to prompts. Gartner places agentic AI at the Peak of Inflated Expectations in 2026.

**Source:** [Gartner — Hype Cycle for Agentic AI 2026](https://www.gartner.com/en/articles/hype-cycle-for-agentic-ai) — accessed 2026-05-07

**See also:** Agent · Agentic firm · Reliability gap

**Compare with deep glossary:** https://agentic-glossary.roei-020.workers.dev/

### Agentic firm

A service company that runs on agents instead of headcount. Operations, compliance, support, marketing delivered by configured agents on a multi-tenant, auditable substrate; humans set policy and review exceptions.

**Source:** [AgentsBooks](https://agentsbooks.com/) — accessed 2026-05-07

**See also:** Agent fleet · Agentic AI · Identity (primitive) · Brain (primitive)

**Compare with deep glossary:** https://agentic-glossary.roei-020.workers.dev/

### Agent fleet

A coordinated set of agents serving one firm or workflow. Fleets share memory, delegate via A2A, and are budgeted, permissioned, and audited as one unit. The fleet, not the individual agent, is the unit of business outcome.

**Source:** [AgentsBooks — Control & Friends](https://agentsbooks.com/features/control-friends) — accessed 2026-05-07

**See also:** Agentic firm · Multi-agent system · A2A / Agent2Agent Protocol · Friends (primitive)

**Compare with deep glossary:** https://agentic-glossary.roei-020.workers.dev/

### Workflow (vs agent)

A system where LLMs and tools are orchestrated through predefined code paths — the path is fixed in advance. Contrast with an agent, which decides the path at runtime. Most production AI systems in 2026 are workflows, not agents.

**Source:** [Anthropic — Building Effective Agents](https://www.anthropic.com/research/building-effective-agents) — accessed 2026-05-07

**See also:** Agent · Agentic AI · Plan-and-execute

**Compare with deep glossary:** https://agentic-glossary.roei-020.workers.dev/

### Tool use / Function calling

*Also known as:* function calling · tool calling

An LLM's ability to call external functions (APIs, databases, file ops) by emitting a structured request the runtime intercepts and executes. The capability that turns a language model into an actor.

**Source:** [OpenAI — Function calling docs](https://platform.openai.com/docs/guides/function-calling) — accessed 2026-05-07

**See also:** Agent · MCP / Model Context Protocol · Brain (primitive) · Toolformer

**Compare with deep glossary:** https://agentic-glossary.roei-020.workers.dev/

### Autonomy

Degree to which an agent operates without human oversight on the path or the result. Anthropic's autonomy research notes that the same property that makes agents useful introduces a range of new risks.

**Source:** [Anthropic — Measuring AI agent autonomy](https://www.anthropic.com/research/measuring-agent-autonomy) — accessed 2026-05-07

**See also:** Agent · Human-in-the-loop (HITL) · Guardrail · Reliability gap

### Multi-agent system

*Also known as:* MAS

A system in which two or more agents collaborate on a task — sharing context, delegating sub-tasks, debating an answer. Multi-agent patterns fan a single user request across multiple specialists and chain results back.

**Source:** [AgentsBooks](https://agentsbooks.com/) — accessed 2026-05-07

**See also:** A2A / Agent2Agent Protocol · Agent fleet · LangGraph · CrewAI · AutoGen

**Compare with deep glossary:** https://agentic-glossary.roei-020.workers.dev/

### AI assistant

An LLM-driven conversational interface that helps a user complete tasks within a host application. Distinct from an agent — assistants typically require a human-in-the-loop turn, while agents act autonomously.

**Source:** [Anthropic — Building Effective Agents](https://www.anthropic.com/research/building-effective-agents) — accessed 2026-05-07

**See also:** Agent · Chatbot · Copilot

### Chatbot

A stateless or near-stateless conversational interface that responds to user prompts. The pre-agent baseline; what 'tools that talk' refers to in the chatbot-vs-agent distinction Gartner draws.

**Source:** [Gartner — Hype Cycle for Agentic AI 2026](https://www.gartner.com/en/articles/hype-cycle-for-agentic-ai) — accessed 2026-05-07

**See also:** Agent · AI assistant

### Copilot

A category of in-context AI assistants embedded in a host application (IDE, office suite, browser) that suggest, complete, or execute actions on behalf of the user. Microsoft's branded usage popularized the term; now generic.

**Source:** [Anthropic — Building Effective Agents](https://www.anthropic.com/research/building-effective-agents) — accessed 2026-05-07

**See also:** AI assistant · Agent

### Generative AI — In force

*Also known as:* GenAI

Class of AI systems that produce novel outputs (text, image, audio, video, code) rather than only classifying or predicting. The substrate underneath all current agentic systems.

**Source:** [NIST — Generative AI Profile (AI 600-1)](https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf) — accessed 2026-05-07

**See also:** Agent · Transformer · Large Language Model (LLM)

### Large Language Model (LLM) — Foundational

*Also known as:* large language model

A neural network trained on broad text (and increasingly code, image, audio) corpora at scale, producing token-by-token generations conditioned on a prompt. The cognitive layer of every agent's brain.

**Source:** [Vaswani et al. — Attention is All You Need (2017)](https://arxiv.org/abs/1706.03762) — accessed 2026-05-07

**See also:** Transformer · Generative AI · Brain (primitive)

### Transformer — Foundational

Neural-network architecture introduced by Vaswani et al. (2017) using self-attention as its sole core mechanism. The substrate of GPT, Claude, Gemini, Llama, and almost every production LLM in 2026.

**Source:** [Vaswani et al. — Attention is All You Need (2017)](https://arxiv.org/abs/1706.03762) — accessed 2026-05-07

**See also:** Large Language Model (LLM) · Generative AI

---

## Protocols

### MCP / Model Context Protocol — In force

*Also known as:* Model Context Protocol · MCP

Open standard defining how AI systems integrate with external data sources and tools. Anthropic introduced MCP in November 2024 and donated it to the Linux Foundation in December 2025; 97M monthly SDK downloads, 10,000+ active servers as of late 2025.

**Source:** [Model Context Protocol — Specification (2025-11-25)](https://modelcontextprotocol.io/specification/2025-11-25) — accessed 2026-05-07

**See also:** A2A / Agent2Agent Protocol · Tool use / Function calling · MCP Tools · MCP Resources · Brain (primitive)

**Compare with deep glossary:** https://agentic-glossary.roei-020.workers.dev/

### MCP Tools — In force

MCP server primitive exposing executable functions the host's LLM can invoke. One of the three core MCP server primitives alongside Resources and Prompts.

**Source:** [Model Context Protocol — Specification](https://modelcontextprotocol.io/specification/2025-11-25) — accessed 2026-05-07

**See also:** MCP / Model Context Protocol · Tool use / Function calling · MCP Resources · MCP Prompts

### MCP Resources — In force

MCP server primitive exposing read-only data the host can pass into the model's context (files, database rows, API responses). Application-controlled, not model-controlled.

**Source:** [Model Context Protocol — Specification](https://modelcontextprotocol.io/specification/2025-11-25) — accessed 2026-05-07

**See also:** MCP / Model Context Protocol · MCP Tools · Knowledge (primitive)

### MCP Prompts — In force

MCP server primitive exposing reusable prompt templates the host can offer to the user as slash commands or quick actions. User-controlled, parameterizable.

**Source:** [Model Context Protocol — Specification](https://modelcontextprotocol.io/specification/2025-11-25) — accessed 2026-05-07

**See also:** MCP / Model Context Protocol · MCP Tools

### MCP Sampling — In force

MCP capability allowing servers to request LLM completions back through the host — letting servers perform agentic behavior without bringing their own model. The reverse direction of normal tool calls.

**Source:** [Model Context Protocol — Specification](https://modelcontextprotocol.io/specification/2025-11-25) — accessed 2026-05-07

**See also:** MCP / Model Context Protocol

### A2A / Agent2Agent Protocol — In force

*Also known as:* Agent2Agent · A2A

Open standard for communication and interoperability between AI agents built by different vendors and frameworks. Originally developed by Google, donated to the Linux Foundation in June 2025; 150+ supporting organizations as of April 2026.

**Source:** [A2A Protocol — Specification](https://a2a-protocol.org/latest/specification/) — accessed 2026-05-07

**See also:** MCP / Model Context Protocol · Agent Card (A2A) · Multi-agent system · Friends (primitive) · A2A Task

**Compare with deep glossary:** https://agentic-glossary.roei-020.workers.dev/

### Agent Card (A2A)

JSON document advertising an A2A agent's capabilities, endpoints, and security requirements. Lets a client agent identify which remote agent can perform a task before delegating. The 2026 v0.3 spec adds signed Agent Cards.

**Source:** [Google Developers — Announcing A2A](https://developers.googleblog.com/en/a2a-a-new-era-of-agent-interoperability/) — accessed 2026-05-07

**See also:** A2A / Agent2Agent Protocol · Multi-agent system

**Compare with deep glossary:** https://agentic-glossary.roei-020.workers.dev/

### A2A Task — In force

Stateful unit of work exchanged between A2A agents, with a defined lifecycle (submitted, working, input-required, completed, failed, canceled). The atom of inter-agent collaboration.

**Source:** [A2A Protocol — Specification](https://a2a-protocol.org/latest/specification/) — accessed 2026-05-07

**See also:** A2A / Agent2Agent Protocol · Agent Card (A2A) · A2A Artifact

### A2A Artifact — In force

Persisted output produced as a result of an A2A task — file, structured data, or message — that downstream agents and the original requester can reference.

**Source:** [A2A Protocol — Specification](https://a2a-protocol.org/latest/specification/) — accessed 2026-05-07

**See also:** A2A Task · A2A / Agent2Agent Protocol

### OpenAI Assistants API

OpenAI's stateful API for building agents with persistent threads, tools, and file search. Marked for deprecation in favor of the Responses API; both surfaces remain available through 2026.

**Source:** [OpenAI — Function calling docs](https://platform.openai.com/docs/guides/function-calling) — accessed 2026-05-07

**See also:** Tool use / Function calling · OpenAI Responses API

### OpenAI Responses API

OpenAI's successor to the Assistants API: a unified interface for tool use, multi-turn agentic loops, and built-in tools (web search, file search, computer use). Released 2025; positioned as the recommended agent surface.

**Source:** [OpenAI — Function calling docs](https://platform.openai.com/docs/guides/function-calling) — accessed 2026-05-07

**See also:** OpenAI Assistants API · Tool use / Function calling

---

## Memory & RAG

### RAG / Retrieval-Augmented Generation — Foundational

*Also known as:* RAG

Architecture combining a generative model with an external retrieval step over a corpus or vector store, so generation is grounded in indexed documents rather than only model parameters. Coined by Lewis et al. (2020); the default pattern for current/proprietary knowledge.

**Source:** [Lewis et al. — Retrieval-Augmented Generation (2020)](https://arxiv.org/abs/2005.11401) — accessed 2026-05-07

**See also:** Memory (primitive) · Knowledge (primitive) · Context graph · Embedding · Vector database

**Compare with deep glossary:** https://agentic-glossary.roei-020.workers.dev/

### Vector database

*Also known as:* vector DB · vector store

Database optimized for similarity search over high-dimensional embeddings. Anchors the retrieval step in RAG and the long-term memory in agent runtimes. 2026 production landscape: Pinecone, pgvector, Qdrant, Weaviate, Milvus, Chroma, Vertex Vector, Vespa.

**Source:** [Vector database landscape — comparison surveys 2026](https://www.firecrawl.dev/blog/best-vector-databases) — accessed 2026-05-07

**See also:** RAG / Retrieval-Augmented Generation · Embedding · Memory (primitive) · Semantic search

### Embedding — Foundational

Dense vector representation of a text (or image, audio) chunk, produced by an embedding model so semantic similarity can be computed via vector distance. The numerical substrate of all retrieval pipelines.

**Source:** [Lewis et al. — Retrieval-Augmented Generation (2020)](https://arxiv.org/abs/2005.11401) — accessed 2026-05-07

**See also:** RAG / Retrieval-Augmented Generation · Vector database · Semantic search · Chunking

### Chunking

Splitting source documents into retrieval-sized fragments before embedding. Strategy choice (fixed-size, sentence, semantic, recursive) significantly affects RAG quality.

**Source:** [Lewis et al. — Retrieval-Augmented Generation (2020)](https://arxiv.org/abs/2005.11401) — accessed 2026-05-07

**See also:** RAG / Retrieval-Augmented Generation · Embedding

### Semantic search

Retrieval over embedding similarity rather than keyword overlap. Powered by vector databases and embedding models; the pure complement to keyword/lexical search.

**Source:** [Lewis et al. — Retrieval-Augmented Generation (2020)](https://arxiv.org/abs/2005.11401) — accessed 2026-05-07

**See also:** Embedding · Vector database · Hybrid search

### Hybrid search

Retrieval combining lexical (BM25/keyword) and semantic (vector) scores, typically reranked together. Outperforms either approach alone for most production RAG use-cases in 2026.

**Source:** [Lewis et al. — Retrieval-Augmented Generation (2020)](https://arxiv.org/abs/2005.11401) — accessed 2026-05-07

**See also:** Semantic search · RAG / Retrieval-Augmented Generation · Embedding

### Context window

Maximum tokens an LLM can attend to in one inference pass. Modern frontier models offer 200K–2M token windows; effective context (where the model actually uses information well) is often shorter than nominal context.

**Source:** [Anthropic — Many-shot jailbreaking](https://www.anthropic.com/research/many-shot-jailbreaking) — accessed 2026-05-07

**See also:** Context graph · Long-term memory · Many-shot jailbreaking

### Context graph — Emerging 2026

Structured representation of entities, relationships, and recent events an agent should keep in working context — distinct from raw vector retrieval. Gartner's 2026 Hype Cycle surfaces context graphs as core engineering practice.

**Source:** [Gartner — Hype Cycle for Agentic AI 2026](https://www.gartner.com/en/articles/hype-cycle-for-agentic-ai) — accessed 2026-05-07

**See also:** Memory (primitive) · RAG / Retrieval-Augmented Generation · Knowledge (primitive) · AX (Agent Experience)

**Compare with deep glossary:** https://agentic-glossary.roei-020.workers.dev/

### Long-term memory

Persistent state an agent retains across sessions — facts, preferences, prior outcomes — typically stored in a database or vector store and retrieved on demand. Distinct from the model's context window.

**Source:** [Park et al. — Generative Agents (2023)](https://arxiv.org/abs/2304.03442) — accessed 2026-05-07

**See also:** Memory (primitive) · RAG / Retrieval-Augmented Generation · Episodic memory · Semantic memory

### Episodic memory

Memory of specific past events the agent participated in — interactions, observations, outcomes — stored as time-stamped records the agent can retrieve and reflect on. Generative Agents (Park et al., 2023) introduced the canonical agent-side framing.

**Source:** [Park et al. — Generative Agents (2023)](https://arxiv.org/abs/2304.03442) — accessed 2026-05-07

**See also:** Long-term memory · Memory (primitive) · Reflexion

### Semantic memory

Memory of general facts, concepts, and procedures the agent has learned — not tied to specific past episodes. Typically realized via knowledge bases, embeddings of documentation, or fine-tuned weights.

**Source:** [Park et al. — Generative Agents (2023)](https://arxiv.org/abs/2304.03442) — accessed 2026-05-07

**See also:** Long-term memory · Knowledge (primitive) · RAG / Retrieval-Augmented Generation

---

## Reasoning

### Chain-of-Thought (CoT) — Foundational

*Also known as:* CoT

Prompting technique that elicits intermediate reasoning steps before a final answer. Wei et al. (2022) showed CoT substantially improves arithmetic, commonsense, and symbolic reasoning. The substrate of most modern reasoning prompts.

**Source:** [Wei et al. — Chain-of-Thought (2022)](https://arxiv.org/abs/2201.11903) — accessed 2026-05-07

**See also:** ReAct · Tree of Thoughts (ToT) · Self-consistency

### ReAct — Foundational

Reasoning-and-acting prompting framework (Yao et al., 2022) that interleaves thought, action, and observation steps. The pattern most agent loops are built on: think, call tool, observe, think again.

**Source:** [Yao et al. — ReAct (2022)](https://arxiv.org/abs/2210.03629) — accessed 2026-05-07

**See also:** Chain-of-Thought (CoT) · Tool use / Function calling · Agent · Plan-and-execute

### Tree of Thoughts (ToT)

*Also known as:* ToT

Reasoning framework (Yao et al., 2023) that explores multiple reasoning paths as a tree, with backtracking and lookahead. Outperforms linear CoT on tasks requiring strategic search.

**Source:** [Yao et al. — Tree of Thoughts (2023)](https://arxiv.org/abs/2305.10601) — accessed 2026-05-07

**See also:** Chain-of-Thought (CoT) · ReAct

### Reflexion

Verbal-reinforcement-learning framework (Shinn et al., 2023) where an agent reflects on prior failures in natural language and stores those reflections as episodic memory for future attempts.

**Source:** [Shinn et al. — Reflexion (2023)](https://arxiv.org/abs/2303.11366) — accessed 2026-05-07

**See also:** Episodic memory · ReAct · Self-consistency

### Self-consistency

Decoding strategy that samples multiple reasoning paths and selects the most consistent answer (typically by majority vote). Improves CoT performance with minimal overhead beyond extra samples.

**Source:** [Wei et al. — Chain-of-Thought (2022)](https://arxiv.org/abs/2201.11903) — accessed 2026-05-07

**See also:** Chain-of-Thought (CoT) · Tree of Thoughts (ToT)

### Toolformer

Self-supervised approach (Schick et al., 2023) that teaches a language model when and how to call APIs by inserting tool calls into training data the model itself selects. Foundational reference for tool-augmented LMs.

**Source:** [Schick et al. — Toolformer (2023)](https://arxiv.org/abs/2302.04761) — accessed 2026-05-07

**See also:** Tool use / Function calling · ReAct

### Plan-and-execute

Agent pattern that produces a multi-step plan first, then executes each step (often with a separate executor model). Tradeoff vs ReAct: more deliberation, less reactivity.

**Source:** [Anthropic — Building Effective Agents](https://www.anthropic.com/research/building-effective-agents) — accessed 2026-05-07

**See also:** ReAct · Workflow (vs agent) · Agent

### Routing

Pattern in which an initial classification step picks which downstream model, prompt, or tool handles the request. Reduces cost and improves quality when subtasks differ in complexity.

**Source:** [Anthropic — Building Effective Agents](https://www.anthropic.com/research/building-effective-agents) — accessed 2026-05-07

**See also:** Workflow (vs agent) · Agent

---

## Frameworks

### LangGraph

Open-source framework (LangChain team) for stateful, durable multi-agent systems via a directed graph of nodes and edges. Most production-mature of the three dominant 2026 agent frameworks; steepest learning curve for finest-grained control.

**Source:** [Independent comparison surveys, 2026](https://pecollective.com/blog/ai-agent-frameworks-compared/) — accessed 2026-05-07

**See also:** CrewAI · AutoGen · LangChain · Multi-agent system

**Compare with deep glossary:** https://agentic-glossary.roei-020.workers.dev/

### CrewAI

Open-source framework modeling multi-agent systems as a team of role-based agents (researcher, writer, reviewer) with shared goals. Easiest learning curve of the three dominant 2026 frameworks; best-fit for business-workflow automation.

**Source:** [Independent comparison surveys, 2026](https://pecollective.com/blog/ai-agent-frameworks-compared/) — accessed 2026-05-07

**See also:** LangGraph · AutoGen · Multi-agent system

**Compare with deep glossary:** https://agentic-glossary.roei-020.workers.dev/

### AutoGen

Microsoft Research multi-agent framework using a conversational group-chat metaphor — agents coordinate by exchanging messages rather than via graph or role hierarchy. Best-fit for debate/vote/iterative-refinement systems.

**Source:** [Independent comparison surveys, 2026](https://pecollective.com/blog/ai-agent-frameworks-compared/) — accessed 2026-05-07

**See also:** LangGraph · CrewAI · Multi-agent system

**Compare with deep glossary:** https://agentic-glossary.roei-020.workers.dev/

### LangChain

Original Python/JS framework for composing LLM applications via chains, retrievers, and tool wrappers. Foundational layer many agent frameworks (including LangGraph) build on; broad ecosystem of integrations.

**Source:** [Independent comparison surveys, 2026](https://pecollective.com/blog/ai-agent-frameworks-compared/) — accessed 2026-05-07

**See also:** LangGraph · RAG / Retrieval-Augmented Generation

### LlamaIndex

Framework focused on data-grounded LLM applications: document ingestion, indexing, retrieval, and query engines. Strong RAG ergonomics; increasingly an agent framework as well in 2026.

**Source:** [Independent comparison surveys, 2026](https://pecollective.com/blog/ai-agent-frameworks-compared/) — accessed 2026-05-07

**See also:** RAG / Retrieval-Augmented Generation · Vector database · LangChain

### Pydantic AI

Type-safe Python agent framework from the Pydantic team emphasizing structured outputs, dependency injection, and testability. Newer entrant in 2024–2026; gaining traction for production-grade agents.

**Source:** [Independent comparison surveys, 2026](https://pecollective.com/blog/ai-agent-frameworks-compared/) — accessed 2026-05-07

**See also:** LangChain · LangGraph

---

## Evaluation

### Eval

*Also known as:* evaluation · evals

A test (or test suite) that measures an LLM or agent's performance on a defined task with grading criteria. Evals are how you tell whether a model change made things better, worse, or neither.

**Source:** [OpenAI — Evals framework](https://github.com/openai/evals) — accessed 2026-05-07

**See also:** Benchmark · LLM-as-judge · HumanEval · SWE-bench · GAIA

### Benchmark

A standardized eval suite published by researchers or vendors so different systems can be compared on the same task definition. HumanEval, SWE-bench, GAIA, MMLU, MMMU are common 2026 references.

**Source:** [OpenAI — Evals framework](https://github.com/openai/evals) — accessed 2026-05-07

**See also:** Eval · HumanEval · SWE-bench · GAIA

### HumanEval — Foundational

Code-generation benchmark from OpenAI's Codex paper (Chen et al., 2021): 164 hand-written Python problems, each scored by passing tests. Long-running standard for measuring LLM coding ability.

**Source:** [Chen et al. — Evaluating Large Language Models Trained on Code (2021)](https://arxiv.org/abs/2107.03374) — accessed 2026-05-07

**See also:** Benchmark · Eval

### SWE-bench

Benchmark (Jimenez et al., 2023) measuring whether agents can resolve real GitHub issues from open-source Python repositories — file-edit-test-fix loops, not isolated functions. The agentic counterpart to HumanEval.

**Source:** [Jimenez et al. — SWE-bench (2023)](https://arxiv.org/abs/2310.06770) — accessed 2026-05-07

**See also:** Benchmark · HumanEval · Agent

### GAIA

General AI Assistants benchmark (Mialon et al., 2023): real-world questions requiring multi-step reasoning, tool use, and web research. Designed to be easy for humans, hard for current agents.

**Source:** [Mialon et al. — GAIA (2023)](https://arxiv.org/abs/2311.12983) — accessed 2026-05-07

**See also:** Benchmark · Agent · Tool use / Function calling

### LLM-as-judge

Grading pattern where a (usually stronger) LLM scores another LLM's output against a rubric. Cheaper and faster than human grading; introduces its own biases (length bias, position bias) that need calibration.

**Source:** [OpenAI — Evals framework](https://github.com/openai/evals) — accessed 2026-05-07

**See also:** Eval · Benchmark

### Hallucination — In force

Confidently-stated output not supported by the model's grounding (training data, retrieved context, or tool results). The dominant reliability problem motivating RAG, citations, and verification harnesses.

**Source:** [NIST — Generative AI Profile (AI 600-1)](https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf) — accessed 2026-05-07

**See also:** Reliability gap · RAG / Retrieval-Augmented Generation · Guardrail

### Reliability gap — Contested

Observed gap between agentic AI's adoption hype and measurable production reliability. McKinsey: 23% of organizations scaling agentic AI but only 39% report enterprise-level EBIT impact. Whether the gap is closing or widening is contested.

**Source:** [Gary Marcus — Six (or seven) predictions for AI 2026](https://garymarcus.substack.com/p/six-or-seven-predictions-for-ai-2026) — accessed 2026-05-07

**See also:** Agentic AI · AI RMF (NIST) · Hallucination

**Compare with deep glossary:** https://agentic-glossary.roei-020.workers.dev/

### Observability

The discipline of making agent runtime behavior visible through traces, spans, metrics, and logs so failures can be diagnosed and improvements measured. 2026 vendor landscape: LangSmith, Langfuse, Arize Phoenix, Helicone, Datadog LLM.

**Source:** [LangChain — LLM observability tools](https://www.langchain.com/articles/llm-observability-tools) — accessed 2026-05-07

**See also:** Trace · Span · Eval · Agent management platform

### Trace

Recorded sequence of all calls (LLM, tool, retrieval, sub-agent) made during a single agent invocation. The atomic unit of agent observability.

**Source:** [LangChain — LLM observability tools](https://www.langchain.com/articles/llm-observability-tools) — accessed 2026-05-07

**See also:** Observability · Span

### Span

Single unit of work within a trace — one LLM call, one tool invocation, one retrieval — with start/end timestamps and metadata. Spans nest hierarchically; OpenTelemetry-compatible spans are the 2026 default schema.

**Source:** [LangChain — LLM observability tools](https://www.langchain.com/articles/llm-observability-tools) — accessed 2026-05-07

**See also:** Trace · Observability

---

## Compliance

### AI RMF (NIST) — In force

*Also known as:* AI RMF · NIST AI 100-1

U.S. NIST AI Risk Management Framework. Voluntary, but widely adopted as the de-facto operating guide for trustworthy AI. Built around four core functions: Govern, Map, Measure, Manage.

**Source:** [NIST — AI Risk Management Framework (AI 100-1)](https://nvlpubs.nist.gov/nistpubs/ai/nist.ai.100-1.pdf) — accessed 2026-05-07

**See also:** GenAI Profile (NIST AI 600-1) · High-risk AI system (EU AI Act) · AI Agent Interoperability Profile (NIST) · ISO/IEC 42001

**Compare with deep glossary:** https://agentic-glossary.roei-020.workers.dev/

### GenAI Profile (NIST AI 600-1) — In force

NIST profile of the AI RMF specific to generative AI, enumerating 12 risks of generative AI (CBRN information, confabulation, dangerous content, data privacy, environmental, harmful bias, human-AI configuration, info integrity, info security, IP, obscene content, value chain) with suggested actions.

**Source:** [NIST — Generative AI Profile (AI 600-1)](https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf) — accessed 2026-05-07

**See also:** AI RMF (NIST) · Trustworthy AI · Hallucination

### High-risk AI system (EU AI Act) — In force 2026-08-02 (subject to deferral)

EU AI Act Annex III categories (critical infrastructure, education, employment, essential services, law enforcement, migration, justice). Subject to conformity assessment, transparency, oversight, robustness, cybersecurity. Originally enforceable 2 August 2026; Digital Omnibus (Nov 2025) proposes deferral to 2 December 2027 — pending trilogue.

**Source:** [European Commission — AI Act Implementation Timeline](https://ai-act-service-desk.ec.europa.eu/en/ai-act/timeline/timeline-implementation-eu-ai-act) — accessed 2026-05-07

**See also:** AI RMF (NIST) · AI Agent Interoperability Profile (NIST) · GPAI (General-purpose AI) · AI literacy (EU AI Act Art. 4)

**Compare with deep glossary:** https://agentic-glossary.roei-020.workers.dev/

### GPAI (General-purpose AI) — In force

*Also known as:* GPAI

EU AI Act category for AI models with significant generality across many tasks (most foundation models qualify). GPAI providers carry transparency, copyright, and — for systemic-risk GPAI — additional obligations.

**Source:** [European Commission — AI Act Implementation Timeline](https://ai-act-service-desk.ec.europa.eu/en/ai-act/timeline/timeline-implementation-eu-ai-act) — accessed 2026-05-07

**See also:** High-risk AI system (EU AI Act) · AI RMF (NIST)

### AI literacy (EU AI Act Art. 4) — In force 2025-02-02

EU AI Act obligation requiring providers and deployers to ensure their staff and others operating AI systems on their behalf have a sufficient level of AI literacy. In effect since 2 February 2025.

**Source:** [European Commission — AI Act Implementation Timeline](https://ai-act-service-desk.ec.europa.eu/en/ai-act/timeline/timeline-implementation-eu-ai-act) — accessed 2026-05-07

**See also:** High-risk AI system (EU AI Act) · GPAI (General-purpose AI)

### AI Agent Interoperability Profile (NIST) — Emerging 2026

Planned NIST profile under the AI RMF specifically targeting agentic systems — covering agent identity, authorization, security, risk management, monitoring, logging. Announced February 2026 via CAISI; planned release Q4 2026.

**Source:** [NIST — AI RMF program](https://www.nist.gov/itl/ai-risk-management-framework) — accessed 2026-05-07

**See also:** AI RMF (NIST) · High-risk AI system (EU AI Act) · A2A / Agent2Agent Protocol

**Compare with deep glossary:** https://agentic-glossary.roei-020.workers.dev/

### ISO/IEC 42001 — In force

International standard (published December 2023) specifying requirements for an AI Management System (AIMS). The first certifiable management-system standard for AI; analogous to ISO 27001 for information security.

**Source:** [ISO — ISO/IEC 42001:2023](https://www.iso.org/standard/81230.html) — accessed 2026-05-07

**See also:** AI RMF (NIST) · Trustworthy AI · High-risk AI system (EU AI Act)

### Trustworthy AI — In force

NIST's umbrella term for AI that is valid and reliable, safe, secure and resilient, accountable and transparent, explainable and interpretable, privacy-enhanced, and fair with harmful bias managed. The seven characteristics framework underneath the AI RMF.

**Source:** [NIST — AI Risk Management Framework (AI 100-1)](https://nvlpubs.nist.gov/nistpubs/ai/nist.ai.100-1.pdf) — accessed 2026-05-07

**See also:** AI RMF (NIST) · ISO/IEC 42001

### Constitutional AI (CAI) — Foundational

*Also known as:* CAI

Anthropic training method (Bai et al., 2022) that gives an AI a written constitution and uses the model itself to critique and revise outputs against those principles, replacing most human-labelled harmlessness data with AI-generated feedback (RLAIF).

**Source:** [Bai et al. — Constitutional AI (2022)](https://arxiv.org/abs/2212.08073) — accessed 2026-05-07

**See also:** Red-teaming · Trustworthy AI · Brain (primitive)

**Compare with deep glossary:** https://agentic-glossary.roei-020.workers.dev/

### Red-teaming — In force

Adversarial testing of an AI system by trying to elicit unsafe, biased, or out-of-policy outputs. NIST AI RMF and EU AI Act both treat red-teaming as a core risk-management practice.

**Source:** [NIST — AI Risk Management Framework (AI 100-1)](https://nvlpubs.nist.gov/nistpubs/ai/nist.ai.100-1.pdf) — accessed 2026-05-07

**See also:** AI RMF (NIST) · Constitutional AI (CAI) · Many-shot jailbreaking · Prompt injection

### Many-shot jailbreaking

Attack pattern (Anthropic, 2024) exploiting long context windows by stuffing the prompt with many examples of unsafe Q&A pairs, conditioning the model to comply with a final unsafe request. Surface area scales with context length.

**Source:** [Anthropic — Many-shot jailbreaking](https://www.anthropic.com/research/many-shot-jailbreaking) — accessed 2026-05-07

**See also:** Red-teaming · Context window · Prompt injection

### Prompt injection — In force

Attack where an attacker controls part of an LLM's input and overrides developer instructions — direct (in the user prompt) or indirect (in retrieved documents, tool outputs, or web pages). The dominant agent-runtime threat vector.

**Source:** [NIST — Generative AI Profile (AI 600-1)](https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf) — accessed 2026-05-07

**See also:** Red-teaming · Guardrail · Sandboxing · Many-shot jailbreaking

### Sleeper agent

Anthropic's term (Hubinger et al., 2024) for a model trained to behave normally except when a specific trigger appears, at which point it executes hidden behavior. Demonstrated that standard safety training can fail to remove such backdoors.

**Source:** [Anthropic — Sleeper Agents (2024)](https://arxiv.org/abs/2401.05566) — accessed 2026-05-07

**See also:** Red-teaming · Constitutional AI (CAI)

### Sandboxing — In force

Containment of agent tool execution inside an isolated environment with constrained filesystem, network, and resource access. The first defensive layer against prompt-injection-driven harmful actions.

**Source:** [NIST — AI Risk Management Framework (AI 100-1)](https://nvlpubs.nist.gov/nistpubs/ai/nist.ai.100-1.pdf) — accessed 2026-05-07

**See also:** Prompt injection · Guardrail · Human-in-the-loop (HITL)

---

## Operations

### ADLC (Agent Development Life Cycle) — Emerging 2026

*Also known as:* ADLC

End-to-end discipline of designing, building, evaluating, deploying, monitoring, and refreshing agents — agentic counterpart to SDLC. Surfaced by Gartner's 2026 Hype Cycle as a core engineering practice for production agentic AI.

**Source:** [Gartner — Hype Cycle for Agentic AI 2026](https://www.gartner.com/en/articles/hype-cycle-for-agentic-ai) — accessed 2026-05-07

**See also:** Agentic AI · AX (Agent Experience) · Agent management platform · Observability

**Compare with deep glossary:** https://agentic-glossary.roei-020.workers.dev/

### AX (Agent Experience) — Emerging 2026

*Also known as:* AX

Agent-side equivalent of UX — the design discipline concerned with how agents perceive, interpret, and act inside an interface or system designed for them, not for human users. llms.txt, structured data, and .md mirrors are AX-friendly affordances.

**Source:** [Gartner — Hype Cycle for Agentic AI 2026](https://www.gartner.com/en/articles/hype-cycle-for-agentic-ai) — accessed 2026-05-07

**See also:** ADLC (Agent Development Life Cycle) · Context graph

**Compare with deep glossary:** https://agentic-glossary.roei-020.workers.dev/

### Agent management platform — Emerging 2026

Class of software that hosts, configures, observes, and governs fleets of agents across a firm — operational counterpart to identity-and-access-management for human staff. Gartner treats this as a distinct emerging category in 2026.

**Source:** [Gartner — Hype Cycle for Agentic AI 2026](https://www.gartner.com/en/articles/hype-cycle-for-agentic-ai) — accessed 2026-05-07

**See also:** Agent fleet · Agentic firm · ADLC (Agent Development Life Cycle) · Observability

**Compare with deep glossary:** https://agentic-glossary.roei-020.workers.dev/

### Guardrail — In force

Runtime check (input filter, output filter, policy-violation detector) that constrains an agent's behavior independent of the model's own reasoning. Used to enforce safety, cost, and scope policies on top of constitutional/aligned models.

**Source:** [NIST — AI Risk Management Framework (AI 100-1)](https://nvlpubs.nist.gov/nistpubs/ai/nist.ai.100-1.pdf) — accessed 2026-05-07

**See also:** Sandboxing · Prompt injection · Human-in-the-loop (HITL)

### Human-in-the-loop (HITL) — In force

*Also known as:* HITL

Operational pattern where humans approve, review, or intervene at defined points in an agent's loop — typically for irreversible or high-stakes actions. NIST AI RMF treats human oversight as a core trustworthy-AI characteristic.

**Source:** [NIST — AI Risk Management Framework (AI 100-1)](https://nvlpubs.nist.gov/nistpubs/ai/nist.ai.100-1.pdf) — accessed 2026-05-07

**See also:** Guardrail · Trustworthy AI · Autonomy

---

## CTAs

- Primary: [Try AgentsBooks Free](https://agentsbooks.com/login?returnTo=/onboarding)
- Secondary: [Read the deep Agentic Glossary](https://agentic-glossary.roei-020.workers.dev/)
- Tertiary: [Read the 8 Primitives Pillar](https://agentsbooks.com/blog/eight-primitives-agentic-firm)

*Built by AgentsBooks · 88 terms · refreshed 2026-05-07*