73 terms · Plain English

The agentic AI glossary.

Every term you’ll meet in the wild — defined in one breath each, no jargon defining jargon. Updated October 9, 2026.

1Password for Agents: A platform from 1Password, launched October 22, 2026, that lets personal AI agents sign in to websites with logins from your vault without ever seeing the password. You approve each login on your own device and 1Password fills it in. Muse and Grok Bot come first; Claude Duet and OpenAI dots are listed as coming soon. Explainer: What is 1Password for Agents?
Agent: A language model wrapped in a role, tools, and memory, running in a loop that plans, acts, and checks its work until a goal is met. The unit of delegation in agentic AI. See how agents work.
Agent Compute Engine: A Google Cloud service, announced October 22, 2026, that runs each AI agent in its own hardware-isolated microVM with its own kernel, disk and short-lived identity, keeping credentials outside the machine. It already powers Gemini’s agents; a customer preview opens in the coming months. Explainer: What is Agent Compute Engine?
Agentic AI: AI that pursues goals with minimal supervision rather than answering single prompts — it plans steps, executes them, and verifies results. The full story: What is agentic AI?
Agent loop: The core cycle behind every agent: plan → act → observe → repeat. A chatbot makes one pass; an agent may circle dozens of times before delivering.
AGI (Artificial General Intelligence): Hypothetical AI matching human capability across virtually all cognitive work. Not the same as agentic AI, which exists today and handles specific, bounded tasks.
Anthropic Public Benefit Fund: A fund announced by Anthropic on October 18, 2026 that holds 5% of its shares (subject to stockholder approval), plus up to 2% more for every 50% rise in the company’s value, to pay for public-good work such as help for workers affected by AI. Explainer: What is the Anthropic Public Benefit Fund?
Autonomy levels: The dial between “drafts and waits for approval” and “executes the whole plan unsupervised.” Consumer agents deliberately sit toward the supervised end: the agent does legwork, you keep judgment calls.
Canva Craft: Canva’s first in-house AI image generation model, announced October 22, 2026. Trained on human-designed Canva templates, it plans the layout first and then renders images up to 2K in about 30 aspect ratios. Available to all users in Dream Lab and through the Canva API. Explainer: What is Canva Craft?
Carbon-X: An open AI model from Hugging Face, released October 22, 2026, that reads raw DNA inside a chat and answers questions about it in English. About 14.8 billion parameters, built on Qwen3-14B, Apache 2.0, and not clinically validated. Explainer: What is Carbon-X?
Chain of thought: A model reasoning step by step — often visibly — before answering. Improves accuracy on multi-step problems and makes the reasoning auditable.
Claude: Anthropic’s family of frontier AI models. The June 2026 lineup spans Haiku 4.5, Sonnet 4.6, Opus 4.8, and the flagship Fable 5.
Claude Duet: Anthropic’s proactive AI agent, launched as a research preview on October 21, 2026. Built on Claude Opus 5.5, it checks connected apps every few hours, sends a morning brief, and completes tasks from its own cloud computer, asking approval before sending, paying, signing or deleting. Included in Claude Pro and Max, US waitlist only at launch. Explainer: What is Claude Duet?
Claude Opus 5.5: Anthropic’s AI model released September 22, 2026, the first of the Claude 5.5 family. It costs $4 in / $20 out per million tokens, about 40% less than Opus 5 on typical work, answers over 30% faster, and tops Opus 5 and Fable 5.1 on Anthropic’s coding and computer-use tests. Claude apps, Claude Code, the API, AWS, Google Cloud, Azure and GitHub Copilot. Explainer: What is Claude Opus 5.5?
Claude dynamic workflows: An Anthropic feature where Claude writes its own program to run many agents in parallel and merge their results. Launched in Claude Code on May 28, 2026 and in Claude Managed Agents (beta, up to 1,000 agents per run) on October 9, 2026. Explainer: What are Claude dynamic workflows?
Claude Fable 5: Anthropic’s Mythos-class frontier model, publicly released June 9, 2026 — its most capable generally available model, with safeguards that reroute a small set of high-risk topics. Explainer: What is Claude Fable 5?
Context window: The working memory of a model — everything it can “see” at once: your brief, the conversation, tool results. Agents manage this budget carefully on long tasks.
Custom agent: An agent you define yourself — persona, tone, expertise, standards — rather than using a built-in specialist. In Agentic AI, custom agents are unlimited with PRO.
DeepInfer: DeepSeek’s in-house inference engine, the software that runs its models in production, open-sourced under Apache 2.0 on October 20, 2026. It supports DeepSeek V4.1 Flash on Nvidia SM100 and Huawei Ascend 950 chips. Not to be confused with the hosting company DeepInfra. Explainer: What is DeepInfer?
DeepSeek V4.1 Flash: The fast, low-cost open-weight model from Chinese lab DeepSeek, first released September 10, 2026 and updated to version 1017 on October 17, 2026. Explainer: What is DeepSeek V4.1 Flash?
Dessa: A full-duplex, speech-to-speech AI model from a small research lab at Block, announced October 23, 2026. Block says it was trained only on recordings of human conversation, with no text, and runs locally on a MacBook Pro M4 Max or better. A research preview with a public waitlist. Explainer: What is Dessa AI?
Fine-tuning: Further training a model on specific data to specialize it. Distinct from role prompting, which specializes behavior without changing the model.
Foundation model: A large model trained on broad data that many different products build upon. The “engine” layer beneath agents and apps.
Claude Fable 5.5: The expected next version of Anthropic’s Fable model. As of October 21, 2026 it is unannounced: the name surfaced in a Microsoft Azure model list on October 13 and in a partner-testing leak. Explainer: Claude Fable 5.5 release date
Frontier model: A model at the current capability ceiling — in mid-2026, the class that includes Claude Fable 5. Agents improve disproportionately with frontier upgrades because errors compound across steps.
Gemini 4 Argon: Google’s frontier AI model, announced September 30, 2026 and widely available since October 13 for Google AI Pro and Ultra subscribers and the paid Gemini API. Known for a 1 million token output limit. Explainer: What is Gemini 4 Argon?
Gemini for small business: A set of small business tools in the Gemini app, announced by Google on October 20, 2026: a business landing page, saved brand styles for images, financial charts from spreadsheets and receipts, and connected apps like Shopify, Canva and Xero. Explainer: Gemini for small business
Gemini Launchpads: A program from Google Public Sector and Accenture Federal Services, announced October 19, 2026, that helps US federal agencies build, deploy and track AI agents running on Gemini, with embedded engineers and pre-built HR, finance and supply chain agents. Explainer: What are Gemini Launchpads?
Gemini skills: Saved instructions in the Gemini app, announced by Google on September 30, 2026, that you run by typing / and a name in any chat. You can stack several, and Gemini can apply them on its own. They replace Gems: personal accounts switch in November 2026, Workspace business in March 2027, schools in June 2027. Explainer: What are Gemini skills?
Generative AI: AI that produces content — text, images, code, audio — in response to prompts. The engine inside agentic systems, but without the goal-pursuit machinery. Comparison: agentic vs. generative.
GPT-6 Astra: OpenAI’s top-tier AI model, released September 3, 2026, built for long agent jobs like computer use, coding and research. It is the first OpenAI model rated “Critical” for cybersecurity risk, so it ships with extra safeguards. Paid ChatGPT plans and the API ($10 in / $50 out per million tokens). Explainer: What is GPT-6 Astra?
Guardrails: Constraints that keep an agent inside intended behavior: forbidden actions, output standards, step budgets, content policies. The reason “autonomous” doesn’t mean “unsupervised.”
Hallucination: A model stating something false with confidence. Agents reduce it via self-checking and citations, but don’t eliminate it — verify before acting on important claims.
Inference: Running a trained model to produce output (as opposed to training it). What you’re paying for, compute-wise, every time an agent thinks.
Kimi K3.1 Pro: Moonshot AI’s flagship model, released October 23, 2026: a post-trained upgrade of Kimi K3 with a 1-million-token context window, priced at $2 per million input tokens and $10 per million output tokens, with open weights promised by November 4. A Highspeed version runs two to three times faster. Explainer: What is Kimi K3.1 Pro?
Kling 4.0: An AI video model from Kling AI, the Kuaishou spinoff, officially launched October 15, 2026. It makes clips up to 30 seconds in one pass, double Kling 3.0, and takes up to 15 reference assets and 10 keyframes through Omni Reference. Explainer: What is Kling 4.0?
LLM (Large Language Model): A neural network trained on vast text corpora to understand and generate language. The technology underneath every modern agent.
Maker–checker: A two-agent pattern: one produces, the other critiques, the first revises. Cheap insurance for code, contracts, and anything where errors are expensive.
MCP (Model Context Protocol): An open standard for connecting AI models to tools and data sources — a universal adapter so any agent can plug into any system without custom wiring.
MCP Tasks: A part of the Model Context Protocol that lets a tool server answer a slow request with a task ID instead of a final result, so the AI keeps working and collects the result later. In public beta in the Claude Code CLI from version 2.1.304 (October 2026). Explainer: What are MCP Tasks?
Memory: What an agent retains: short-term (the running task context) and sometimes long-term (durable notes across sessions). Keeps step fourteen consistent with step two.
Mercury 3: Inception Labs’ diffusion language model, released October 19, 2026. It drafts many tokens at once instead of one by one, reaching about 1,068 tokens per second, at $0.20 per million input tokens. Explainer: What is Mercury 3?
Microsoft Decision-1: A decision model from Microsoft, announced October 9, 2026. Built on Alibaba’s Qwen3.5-9B, it returns a calibrated probability for each of a fixed set of answers instead of writing text, for routing, classification and checking agent actions. Explainer: What is Microsoft Decision-1?
MiniMax M3.1 Flash: A fast, low-cost AI model for agent tasks from Chinese company MiniMax, launched October 19, 2026 at $0.15 per million input tokens. It was tested anonymously as "Space Bunny." Explainer: What is MiniMax M3.1 Flash?
Mistral Large 4: Mistral’s flagship open-weight model, nicknamed “Le Chonk,” in public preview since October 6, 2026. It has about 1 trillion parameters (about 49B active per token), reads text and images, and its weights are due at the end of October 2026. Explainer: What is Mistral Large 4?
Multi-agent system: Multiple specialized agents sharing one session, reading each other’s output, and dividing labor — research, strategy, writing, planning. Why it wins: two agents beat one.
MultiLangBench: OpenAI’s own coding evaluation suite, shared October 19, 2026, that asks a model to build a SQLite-compatible database engine in each of 20 programming languages, including COBOL, Fortran and Visual Basic. GPT-6 Astra averaged 87.4%. Explainer: OpenAI’s COBOL update
Multimodal: Able to work across media types — text, images, documents, audio — not just words. Lets agents read screenshots, charts, and PDFs.
Muse Spark Open: Meta’s open-weight AI model, released October 15, 2026: 816 billion parameters (about 42 billion active), a 1 million token context, and a Meta license that lets most people use it for free. Explainer: What is Muse Spark Open?
Muse Wonder: Meta’s first frontier AI models, Muse Wonder and Muse Wonder Full, launched October 12, 2026 for coding and long agent tasks. Muse Wonder is on the Meta Model API; Wonder Full is only in paid Muse Code plans. Explainer: What is Muse Wonder?
Nemotron 3.5: Nvidia’s family of free, open AI models built for agents. Super VL (120 billion parameters, 12 billion active) and Bolt 74B-A9B (fits in 24GB for local AI PCs) launched October 15, 2026. Explainer: What is Nemotron 3.5?
NVIDIA OpenShell: NVIDIA’s free, open-source (Apache 2.0) runtime that runs an AI agent inside a sandbox and enforces policy on its file access, commands and network connections, adding credentials only for approved sites. The OpenShell Reference Architecture, which adds Vera CPUs and the BlueField-4 Sentry monitor, launched October 21, 2026, at GTC Berlin. Explainer: What is NVIDIA OpenShell?
OpenAI Decisions API: An OpenAI developer endpoint, generally available since October 22, 2026, that answers closed questions with a probability, a pick from a fixed list, or a score instead of writing text. It runs on GPT-6 Luna at $0.10 per million input tokens with free output. Explainer: What is the OpenAI Decisions API?
Orchestration: The coordination layer of a multi-agent system: who works when, what each agent sees, how outputs combine. In consumer apps, orchestration ships pre-built.
Panel: A multi-agent pattern where specialists answer the same question from different angles, then reconcile. Suits decisions and evaluations.
PHP2Shell: The nickname for CVE-2026-109618 and CVE-2026-109619, two critical PHP engine flaws (CVSS 9.2) that let an attacker run code on a web server with one unauthenticated request. PHP patched them on October 21, 2026, in versions 8.5.12, 8.4.27, 8.3.36 and 8.2.35. Calif found one of them using OpenAI’s GPT-5.6 Sol. Explainer: What is PHP2Shell?
Pipeline: A multi-agent pattern where each agent builds on the previous one’s output — researcher → strategist → writer → planner. Suits staged work like launches.
Prompt: The instruction you give a model. For agents, the better word is brief: goal, constraints, definition of done. Templates: the prompts playbook.
Prompt engineering: The craft of writing instructions that get reliably good output. With agents, it converges on writing good briefs — the same skill as delegating to people.
RAG (Retrieval-Augmented Generation): Fetching relevant documents and giving them to a model before it answers, so responses are grounded in real sources instead of memory alone.
Reasoning model: A model variant that spends extra computation “thinking” before answering — stronger on math, code, and planning. Most frontier models now blend this in.
Role (system prompt): The standing instructions that define who an agent is and what good output looks like. Roles are behavioral constraints, not costumes — a defined specialist outperforms a vague generalist.
Sandbox Runtime (srt): Anthropic’s free, open-source tool that limits which files and websites an AI agent or any program can reach, using the operating system’s own sandboxing (Seatbelt on macOS, bubblewrap on Linux) instead of a container. Built for Claude Code and open-sourced on October 20, 2025, as a research preview. Explainer: What is Anthropic Sandbox Runtime?
ServiceNow Arc: A personal AI agent from ServiceNow, built with Nvidia, that plans and completes multi-step work across email, Microsoft Teams and ServiceNow under rules set by IT, running inside NVIDIA OpenShell. Previewed as Project Arc in May 2026 and generally available October 21, 2026, on Mac. Explainer: What is ServiceNow Arc?
Spark X2.6: iFlytek’s large language model, launched October 22, 2026 and trained on Chinese chips: 293 billion parameters (30 billion active) with a 1 million token context window. A small open version, Spark-X2.6-4B, is free under Apache 2.0. Explainer: What is Spark X2.6?
Team session: Agentic AI’s multi-agent mode: up to four experts on one brief, each seeing the running conversation and adding only what’s missing. Like a smart meeting, in your pocket.
Timestral: Mistral’s first family of time-series forecasting models, released October 15, 2026, in four sizes from 5M to 209M parameters. It predicts numbers like sales or energy use and runs on a regular CPU. Explainer: What is Timestral?
Token: The unit models read and write — roughly ¾ of an English word. Context windows and API prices are measured in tokens.
Tool use (function calling): A model invoking external capabilities — search, code execution, file reading — instead of just describing them. How agents touch reality; results feed the agent loop’s “observe” step.
Voxtral 3: Mistral’s third generation of speech models, launched October 20, 2026: three speech-to-text models, the Voxtral Speech 3 text-to-speech model and a voice agent SDK built on LiveKit for AI that listens, thinks and talks in real time. Explainer: What is Voxtral 3?
Walmart Agent Gateway: A direct connection, announced by Walmart on October 21, 2026, that lets approved AI agents like ChatGPT, Gemini and Muse search and buy at Walmart over MCP. It comes with Agent Profiles, which give each agent its own spending limit (such as $50 or $100) or require approval for every order, and agents never see your card. Explainer: Walmart AI agents, explained
Workflow vs. agent: A workflow follows a fixed, predefined sequence; an agent chooses its own steps based on what it observes. Workflows are predictable; agents are adaptable. Real products mix both.
x401 protocol: An open web protocol from Proof, launched October 15, 2026, that lets a website ask an AI agent for verified proof of the person it acts for and that person’s signed approval of a specific transaction. It extends HTTP 401 and pairs with x402 for payments. Explainer: What is the x401 protocol?

Agentic AI: Vocabulary is nice; delegation is nicer. Meet your agent team. Download free