Das KI-Agenten-Glossar.

Google's 2025 open Agent2Agent protocol — a standard for agents from different vendors to discover each other, exchange tasks, and stream results.

A RAG variant that routes queries to different retrieval strategies based on complexity — simple questions skip retrieval, hard ones get multi-hop retrieval.

A software system powered by an LLM that perceives its environment, plans actions, and executes them — usually across multiple steps and tools.

The persistent state an agent maintains across turns and sessions — covers short-term context, long-term facts, episodic events, and procedural skills. Distinct from the LLM context window.

Specialized observability for AI agents — tracing the agent's reasoning, tool calls, sub-agent communication, state changes, and decision points across a multi-step run.

The control layer that coordinates multiple agents or agent steps — routing work, managing state, enforcing hand-off rules, and resolving conflicts between specialized agents.

The umbrella term for the standards agents use to communicate with tools (MCP) and with each other (A2A) — the connective-tissue layer of the 2026 agent ecosystem.

An isolated execution environment — usually a container, microVM, or browser profile — where an agent can run code, browse, and act without affecting the host system or shared state.

A multi-environment benchmark suite for LLM-as-agent performance — covers OS, database, web shopping, knowledge graph, card game, and lateral-thinking tasks across 8 environments.

AI systems that act with autonomy — perceiving their environment, planning multi-step actions, calling tools, and iterating toward a goal — as opposed to single-turn generative AI that only responds to prompts.

A web browser where an AI agent is a first-class user — given a goal, the browser plans, navigates, clicks, and fills forms across pages. Arc Search, Dia, and Comet are 2026 examples.

The emerging pattern where an AI agent — not a human — researches, compares, and transacts on behalf of the user. 2026 standards from Visa, Mastercard, and Stripe are formalizing it.

The core control flow of an agent: observe → reason → act → observe, repeated until the goal is met or a stop condition fires.

A retrieval pattern where the agent decides when and what to retrieve — issuing its own search queries, refining them, and iterating — instead of a single up-front retrieval step.

A search pattern where an agent — not a single retrieval call — runs the query: it plans, queries multiple sources, evaluates results, refines, and returns a synthesized answer with citations.

A workflow where an AI agent plans, executes, and adapts a multi-step process — autonomously calling tools, reading data, and looping until the goal is met.

A library or toolkit for building AI agents — providing primitives for tool calling, planning, memory, and orchestration so you do not rebuild the agent loop from scratch.

The research and engineering practice of ensuring AI systems pursue the goals their designers intend — covering training-time techniques like RLHF and constitutional AI as well as deployment-time guardrails.

An AI business development representative — an autonomous sales agent focused on outbound prospecting, top-of-funnel qualification, and pipeline generation, often used interchangeably with AI SDR.

Systematic errors in AI outputs that disadvantage specific groups, perspectives, or topics — caused by biased training data, biased reward signals, or biased evaluation criteria.

AI outputs that include verifiable links or references to source documents — the trust primitive that separates research-grade AI from pure generative chat.

An agent that reviews pull requests — reads the diff, finds bugs, flags style and security issues, and posts inline comments. GitHub Copilot Code Review, CodeRabbit, Greptile, and Cursor BugBot lead the 2026 market.

The classifier and policy layer that filters input to and output from an LLM agent — blocks unsafe categories (CSAM, self-harm, malware), enforces brand voice, and flags PII.

An agent that connects to data warehouses, runs SQL, builds charts, and produces narrative analyses — replacing the "Slack-to-the-analytics-team" loop for routine business questions.

The phenomenon where an AI system's behavior changes over time without explicit code changes — caused by model version updates, training data shifts, or vendor-side changes.

Marketing-grade synonym for a digital worker — an agent positioned as a hireable, role-shaped teammate. Notable 2024–2026 examples: 11x's Alice, Artisan's Ava, Devin from Cognition.

Systematic test suites for AI systems — input/expected-output pairs run automatically to catch regressions when models or prompts change.

A middleware layer that sits between your application and LLM APIs — handles routing, fallback, caching, rate limiting, cost tracking, and observability across multiple model providers.

The framework of policies, controls, and review processes that ensure AI systems are deployed safely, ethically, and in compliance with regulation — covers risk management, audit trails, and stakeholder accountability.

The transition pattern where an AI agent transfers control of a task to a human or to another agent — preserving context, state, and prior actions so the receiver can continue seamlessly.

A staged framework describing organizational AI evolution — typically experimenting → piloting → scaling → optimizing → transforming. Used to plan investment + measure progress.

An agent that joins meetings (or processes recordings) to transcribe, summarize, extract action items, and follow up — Otter, Fireflies, Granola, and Read.ai are 2026 leaders.

OpenAI's browser-based autonomous agent that takes a task ("book me a flight to SFO next Tuesday") and completes it by driving a real web browser — clicking, typing, navigating, and confirming.

A workflow where an engineer codes alongside an AI assistant that suggests, completes, and reviews code in real time — distinct from autonomous coding agents that ship PRs without human intervention.

A time-boxed, scope-limited deployment of an AI agent against a real workflow to measure quality, cost, and adoption before broader rollout. The standard 2026 enterprise procurement pattern.

A multi-step data processing flow that includes one or more LLM or AI calls — typically combines preprocessing, retrieval, LLM inference, post-processing, and observability into a single deployable unit.

An assessment of an organization's preparedness to deploy AI productively — covering data infrastructure, talent, governance, and use-case maturity.

An agent that takes a research question, searches multiple sources over multiple rounds, synthesizes a sourced report, and follows up with clarifications. Distinct from a search box.

The business return generated by an AI deployment minus its full cost — model spend, infra, integration, change management, and risk. The 2026 procurement north-star metric.

The research and engineering discipline focused on making AI systems behave reliably, refuse harmful requests, and fail gracefully under unexpected inputs — covering both training-time alignment and deployment-time guardrails.

An AI sales development representative — an autonomous agent that handles prospecting, outbound email sequencing, and lead qualification end-to-end.

Sending model output to the user token-by-token as it generates, instead of waiting for the full response. The default UX pattern for AI chat in 2026.

Techniques that embed a detectable signal in AI-generated text, images, audio, or video so downstream systems can identify content as machine-generated.

The collective fleet of AI agents and digital workers an organization runs — managed as a unit with shared governance, shared identity, shared observability, and a unified cost model.

The RAG eval metric that scores whether the answer actually addresses the user's question. Catches the "perfectly grounded but useless" failure mode.

François Chollet's benchmark for measuring fluid intelligence — agents must induce a transformation rule from a few input/output grid examples and apply it. Designed to resist memorization.

A UI pattern where AI-generated content (code, documents, diagrams) renders in a separate panel beside the chat, so users can edit and iterate without losing the conversation.

The neural-network primitive that lets a transformer model weigh the importance of every input token when generating each output token — the core innovation behind LLMs.

Microsoft's open-source framework for multi-agent conversation — agents talk to each other to solve problems collaboratively, with explicit support for code execution and human-in-the-loop.

The 2023 open-source project that popularized autonomous LLM agents — wraps an LLM in a recursive plan-execute-reflect loop with persistent goals and tool use.

An agent that plans, executes, and finishes a multi-step task without asking for human approval between steps.

A minimal 2023 reference implementation of an autonomous task-driven agent — three loops (task creation, prioritization, execution) in ~100 lines of Python.

Running model inference asynchronously over a large batch of inputs, traded for latency. OpenAI/Anthropic batch APIs are typically 50% cheaper than sync calls.

AWS's managed agent service inside Amazon Bedrock — provides agent orchestration, tool integration via OpenAPI/Lambda, and a knowledge-base layer for RAG out of the box.

A publicly-shared, standardized eval suite used to compare models and agents across a uniform task — SWE-bench, MMLU, GAIA, etc.

An AI agent specialized in driving a web browser — navigating sites, filling forms, scraping data, and completing multi-step web workflows on behalf of the user.

An agent capability where the LLM drives a real web browser to read, click, and fill forms on live websites.

A deployment pattern where you supply your own model API key to the agent — token costs are billed to you directly, the agent vendor charges only for the software.

OpenAI's side-panel editing surface for documents and code generated by ChatGPT — the OpenAI equivalent of Claude's Artifacts.

An architecture where agents run in sequence, each refining or extending the previous agent's output. Used for long documents or multi-stage workflows.

A prompting technique that asks the model to lay out its reasoning step-by-step before committing to an answer — improves accuracy on multi-step tasks.

An eval metric for systems that cite sources — measures whether citations resolve to real documents, point to the supporting passage, and match the cited claim.

An agent capability for writing and running code in a sandboxed environment — usually Python — to compute, transform data, or test hypotheses.

An AI agent specialized in software engineering tasks — reading codebases, writing code, running tests, opening pull requests, and fixing bugs.

An agent capability where the LLM controls a computer's mouse, keyboard, and screen directly — interpreting screenshots, clicking, typing, and navigating arbitrary desktop and browser apps.

The AI field focused on letting machines understand images and video — covers object detection, image classification, segmentation, OCR, scene understanding, and more.

An alignment technique developed by Anthropic where the model is trained to follow a written set of principles ("a constitution") rather than per-example human preferences — produces safer behavior without massive human-labeling effort.

The discipline of curating what information goes into an LLM's context window — selecting, ordering, and formatting the system prompt, examples, retrieved documents, and conversation history for maximum effectiveness.

The maximum number of tokens a model can consider at once — covers the system prompt, conversation, tool results, and the answer being generated.

An AI tool that suggests changes inline and waits for the user to accept — the human stays in the driver's seat.

A RAG variant that grades retrieved documents and triggers fallback retrieval (web search, alternative sources) when the initial retrieval scores low on relevance.

The fully-loaded cost of an AI completing one unit of work — model spend + infrastructure + integration cost amortized + retries. The right unit for AI ROI math.

An open-source Python framework for role-based multi-agent systems — define agents with roles, goals, and tools, then orchestrate them into "crews" that collaborate on tasks.

An agent capability that produces long-form, multi-source research reports by autonomously browsing the web, reading documents, and synthesizing findings — typically running for 5–30 minutes per query.

In support agents: the percentage of customer contacts the agent resolves fully without escalating to a human.

The standard modern retrieval approach where queries and documents are encoded as dense embedding vectors and matched by similarity — distinct from sparse retrieval (BM25, keyword search).

A persistent agent that occupies a named role within a team — has a job description, KPIs, access to specific tools, and is managed alongside human teammates. The 2026 enterprise framing of agent deployment.

Direct Preference Optimization — a simpler alternative to RLHF that trains models directly on preference data without needing a separate reward model or reinforcement learning loop.

A Stanford-built framework that treats LLM prompts as compilable programs — define what you want declaratively, DSPy optimizes the prompts and few-shot examples automatically.

AI that runs on the device where data is generated — phone, laptop, IoT, vehicle, factory floor — rather than in a remote data center. Trades model size for latency, privacy, and offline operation.

A neural network specialized for converting text (or images, audio) into fixed-length dense vectors — used for semantic search, RAG, clustering, and similarity tasks.

Dense numerical vector representations of text, images, or audio — used to measure semantic similarity, power search, and ground LLM outputs in your data.

Capabilities that appear suddenly above a certain model scale — chain-of-thought reasoning, in-context learning, instruction following — and are absent or near-zero in smaller models.

The agent's memory of specific past events and sessions — "what happened when" — usually stored as timestamped summaries that can be retrieved by time, topic, or participant.

The European Union's regulatory framework for AI systems — categorizes AI by risk level (prohibited, high-risk, limited risk, minimal risk) and imposes obligations based on category. Phased into force 2024–2027.

A systematic test that measures agent performance on a fixed set of inputs — the agent equivalent of a test suite.

The RAG eval metric that scores whether the answer's claims are supported by the retrieved context — the standard RAGAS metric and a near-synonym for groundedness.

A prompting technique where the LLM sees a small number of input/output examples in the prompt before being asked to perform the same task on a new input.

The process of training a pre-trained LLM on additional data to adapt it for a specific task, domain, or style — produces a specialized model derived from a general-purpose base.

A pricing model where the agent has a useful free tier with paid plans for higher usage, more features, or commercial use.

The current generation of state-of-the-art LLMs — typically the largest models from OpenAI, Anthropic, Google, and a small number of others.

An LLM API feature that lets the model emit a structured JSON call to a developer-defined function — the model picks the function name and arguments; the runtime executes the call.

A 466-question benchmark from Meta + Hugging Face that tests general-purpose AI assistants on real-world tasks requiring web browsing, file handling, and multi-step reasoning.

A reasoning structure that generalizes Tree of Thoughts to an arbitrary DAG — intermediate thoughts can be combined, refined, or referenced from multiple branches.

A RAG eval metric measuring whether the generated response is supported by the retrieved context. Distinct from factual accuracy — the answer could be grounded in a wrong source.

Constraints and filters layered around an LLM that prevent it from producing harmful, off-topic, or policy-violating outputs — applied at input, output, or both.

When an LLM generates content that sounds plausible but is factually wrong or fabricated — a citation that doesn't exist, a function that isn't in the API.

A multi-agent architecture where a "manager" or "planner" agent delegates sub-tasks to specialist worker agents — the most common multi-agent pattern in 2026 production systems.

A workflow pattern where the agent pauses for human approval at one or more checkpoints before continuing.

A code-generation benchmark from OpenAI: 164 Python programming problems with unit tests, used to measure an LLM's ability to generate correct code from a natural-language description.

A retrieval technique that combines vector (semantic) search with keyword (lexical) search, fusing the scores to get higher precision than either alone. The 2026 production-grade default for RAG.

The broader category of AI-generated images — includes text-to-image, image-to-image (editing), inpainting, outpainting, and style transfer. Powered by diffusion models or transformer-based image generators.

An LLM's ability to learn a new task at inference time by reading examples in the prompt — no weight updates, just pattern-matching from context.

The process of running a trained LLM to produce outputs — the production phase, distinct from training. Inference is what you pay for when you use an LLM API.

Spending more compute at inference (longer reasoning chains, multiple samples, search) to improve quality on hard problems — the architectural bet of 2025–2026 reasoning models.

A fine-tuning technique where a pre-trained LLM is trained on instruction-response pairs so it learns to follow natural-language commands instead of just predicting next tokens.

A prompting technique that bypasses an LLM's safety guardrails to make it produce content the model was trained to refuse.

An LLM API setting that constrains the output to syntactically valid JSON — and, with strict mode, to a specific schema. The simplest reliable path to structured output.

A transformer inference optimization that stores key/value attention tensors from previous tokens so they do not need to be recomputed on every new token.

The original Python/TypeScript framework for building LLM applications — provides abstractions for chains, agents, tool use, memory, and retrieval. In 2026, mostly superseded by LangGraph for new projects.

An open-source framework from LangChain for building stateful, multi-step agent applications as graphs — nodes are agent steps, edges define control flow, and state persists across steps.

A Python framework focused on RAG and data-augmented LLM applications — provides indexing, retrieval, and query pipelines for connecting LLMs to your data.

An evaluation pattern where a stronger LLM scores another LLM's outputs — replacing or supplementing human review when exact-match grading is infeasible.

A specific kind of AI gateway focused on LLM API calls — provides a unified interface to multiple LLM providers (OpenAI, Anthropic, Google, Mistral, etc.) with routing, caching, and observability.

The practice of monitoring, tracing, and debugging LLM-powered systems in production — capturing prompts, completions, latency, cost, and errors across every call.

A proposed convention (like robots.txt) for sites to tell LLMs which content to ingest, in what summary form, and on what terms. Adoption growing through 2025–2026.

A large language model running entirely on hardware you control — your laptop, your server, or your data center — with no calls to external APIs.

The agent's memory that survives across sessions, sometimes across months or years — usually a vector store plus a key-value store, with episodic, semantic, and procedural layers underneath.

Low-Rank Adaptation — a parameter-efficient fine-tuning technique that updates a small number of additional weights instead of the full model, cutting compute and storage cost by 100×+ with minimal accuracy loss.

A process that exposes tools, resources, or prompts over the Model Context Protocol — any MCP-compliant agent can connect to it and use what it exposes.

The mechanism by which an agent remembers information across sessions — usually a vector store or structured key-value cache.

An architecture where multiple agents (often using different models) generate candidate responses, then an aggregator agent synthesizes them. Higher quality at higher cost.

A model architecture where multiple specialized expert networks share the work — a routing layer activates only a few experts per input, cutting inference cost while keeping total parameter count high.

Massive Multitask Language Understanding — a 57-subject multiple-choice benchmark spanning STEM, humanities, social sciences, law, and ethics. The default measure of "general knowledge" for LLMs since 2020.

A short structured document published with an AI model — declares intended uses, training data overview, performance across subgroups, known limitations, and risk factors.

Model Context Protocol (MCP) is an open standard that lets any AI agent connect to any tool or data source through a single protocol — solving the M×N integration problem for the agent ecosystem.

A training technique that transfers knowledge from a large "teacher" model to a smaller "student" model by training the student to match the teacher's outputs — produces a faster, cheaper model that retains most of the teacher's capability.

A component that selects which LLM to use for each request — based on cost, latency, capability, or content classification. Sits inside an AI gateway or as a standalone routing layer.

The infrastructure layer that hosts a model and exposes inference over HTTP — covering batching, scheduling, KV-cache management, and request routing.

A multi-turn conversation benchmark where models are judged by a strong "LLM-as-judge" on 80 open-ended questions across writing, reasoning, math, coding, and roleplay.

An architecture where several specialized agents collaborate on the same task — each handles a sub-goal and they coordinate through a shared workspace.

The ability of an LLM or agent to chain multiple inferences together to solve a problem — answer A leads to question B, which leads to question C, and so on until the final answer.

AI systems that process and reason across multiple input types — text, images, audio, video — within a single model, instead of routing each modality through separate specialized models.

The AI subfield focused on extracting meaning from human language — intent classification, entity extraction, sentiment analysis, and semantic interpretation. In 2026, mostly subsumed by LLMs.

A computational model loosely inspired by biological neurons — layers of weighted nodes that transform inputs to outputs. LLMs are large neural networks; so are image classifiers, recommendation systems, and most modern AI.

AI tools that let non-engineers build agents, workflows, or applications via visual interfaces — drag-and-drop, prompts, or declarative configuration instead of writing code.

Technology that extracts text from images, scanned documents, and PDFs — in 2026, OCR is often built into multimodal LLMs (Claude, GPT-4o, Gemini) rather than requiring a separate service.

A deployment where the agent runs entirely on infrastructure the customer controls — no agent code or customer data leaves the customer's network.

An agent whose source code is publicly licensed (MIT, Apache, AGPL) — you can self-host, fork, and audit.

OpenAI's 2025 production framework for building agents — successor to the older Assistants API, with first-class handoffs, guardrails, tracing, and tool use.

A pricing model where the vendor charges per successful outcome — closed ticket, qualified lead, resolved bug — rather than per seat, per task, or per token. The signature 2026 agent pricing pattern.

A model capability where the LLM returns multiple tool calls in a single response — the agent runtime executes them concurrently rather than serially, cutting latency on independent operations.

A pricing model where you pay per completed task — per PR generated, per ticket resolved, per email drafted — rather than per seat or per month.

A canonical two-stage agent pattern: a planner LLM produces a structured multi-step plan, then an executor (often a cheaper model) carries out each step using tools.

The phase where an agent decomposes a goal into a structured sequence of sub-tasks before executing any of them.

Running LLM inference inside your security perimeter (VPC, on-prem, confidential compute) so prompts and outputs never leave your control. Mandatory for regulated industries.

A vendor-side optimization that reuses computation for shared prompt prefixes across requests — billed at a 75–90% discount compared to fresh prompt tokens.

The practice of designing, refining, and testing the text instructions sent to an LLM to maximize output quality — covers system prompts, few-shot examples, formatting, and meta-instructions.

An attack where malicious instructions are smuggled into an LLM's input — through user prompts, web pages, documents, or tool outputs — causing the agent to ignore its real instructions.

Parameterized prompt patterns stored as reusable, version-controlled assets — the basic abstraction for managing prompts at production scale.

The practice of treating system prompts as first-class code — versioned, tested, and deployed through CI/CD instead of edited inline in source files.

A Python agent framework from the Pydantic team — type-safe agents with structured outputs, model-agnostic, and a thin API designed to feel like FastAPI for LLMs.

A technique that reduces model weights from 16-bit or 32-bit floats to smaller representations (8-bit, 4-bit, or lower), cutting memory use and inference cost by 2–8× with minimal accuracy loss.

Retrieval-augmented generation — pulling relevant documents from a knowledge base before generating, so the LLM grounds its answer in your data.

An open-source RAG evaluation framework — the de facto standard in 2026 for measuring faithfulness, answer-relevance, context-precision, and context-recall.

An agent built on the ReAct pattern: an interleaved loop of reasoning (the model thinks out loud) and acting (the model calls a tool), repeated until the goal is met.

A class of LLM (o3, Claude Sonnet 4.6, Gemini 2.5 reasoning) that produces a long internal chain of thought before responding — trading latency for accuracy on hard problems.

A structured testing practice where adversaries actively try to break an AI system — finding jailbreaks, hallucinations, harmful outputs, or unsafe tool calls before attackers do.

An agent design pattern where the agent reflects on its previous attempts, generates a critique, and uses the critique to improve subsequent attempts — produces measurable accuracy gains on hard tasks.

A second-stage retrieval model that re-scores a small set of candidate results from initial retrieval — using a cross-encoder or LLM to produce more accurate final rankings.

Reinforcement Learning from Human Feedback — a training technique where humans rate model outputs to teach the model which responses are preferred, dramatically improving instruction-following and safety.

Empirical power-law relationships between model size, training data, and compute that predict loss and capability — the basis for every major frontier-model training plan since 2020.

The classic SaaS pricing model where customers pay per active user — common for copilot-style products (Cursor, GitHub Copilot, Notion AI) but eroding for autonomous agents.

A reasoning technique where the model samples multiple chain-of-thought traces for the same problem and selects the most common final answer — cheap accuracy boost on math and logic tasks.

An agent capability where the model evaluates its own output for errors and produces a corrected version — improves accuracy on verifiable tasks by 10–30% in 2026.

A RAG variant where the model decides on the fly whether to retrieve, what to retrieve, and whether its own draft is grounded — emitting reflection tokens at each step.

An agent capability where the model generates an explicit reflection on its own reasoning or outputs — used to improve subsequent steps or detect errors before they propagate.

A cache layer that matches incoming prompts to past prompts by embedding similarity rather than exact match — serves stored responses for paraphrased queries.

A document-splitting technique that uses embeddings to detect semantic boundaries — produces more coherent chunks for RAG than fixed-size chunking, improving retrieval quality.

The agent's store of timeless facts — "the user is a VP of Sales at Acme," "the company uses Snowflake" — distinct from events (episodic) or skills (procedural).

A routing technique that uses embedding similarity to send each request to the right model, agent, or workflow — instead of brittle keyword rules or expensive LLM classifiers.

Search that ranks results by meaning rather than keyword overlap — using vector embeddings or LLM reasoning to match queries with conceptually similar content.

An agent that plans and executes most steps unsupervised but pauses for approval before anything irreversible.

An LLM inference and programming framework optimized for structured generation, agent workloads, and complex prompting patterns — competitive with vLLM on throughput and faster on JSON/grammar-constrained output.

AI tools that employees use at work without IT or security approval. The 2026 successor to "shadow IT" — broader, faster-spreading, and harder to govern.

A capable LLM in the 1B–13B parameter range — trained to compete with frontier-quality on specific tasks while running on consumer hardware or at fraction-of-frontier cost.

An inference optimization where a small draft model proposes multiple tokens at once and the large model verifies them in parallel — same output, 2–4× faster.

AI technology that converts spoken audio into written text — also called Automatic Speech Recognition (ASR). The input half of voice AI, distinct from TTS which produces speech.

Serving LLM outputs token-by-token as they're generated, typically over SSE or WebSocket — the default deployment pattern for any user-facing AI in 2026.

A model feature that constrains output to a specific JSON schema, making LLM responses safely parseable by downstream code.

A flat-rate monthly price per user — the dominant pricing model for agents aimed at individual contributors.

In hierarchical multi-agent systems, the top-level agent that delegates work to specialist sub-agents, monitors progress, handles failures, and aggregates results.

A multi-agent pattern where many similar agents collaborate without a central supervisor — inspired by ant colonies and bee swarms, used for parallel exploration and consensus.

A benchmark from Princeton that tests coding agents on real GitHub issues — given the bug report and repo, the agent must produce a patch that passes the project's tests.

The initial instruction text given to an LLM that sets its persona, tools, constraints, and default behavior for the session.

The agent reasoning step where a high-level goal is broken into ordered, executable sub-tasks before any tool call is made. Foundational to plan-and-execute, ReAct, and tree-of-thoughts patterns.

Total cost of ownership — the all-in cost of running an agent including subscription, token spend, ops time, and integration work.

The LLM sampling parameter that controls randomness — low values stay near the most-likely next token, high values explore the distribution. 0–2 typical range.

Spending more compute at inference time — longer reasoning, more samples, search — to get higher accuracy without retraining the model.

AI technology that generates images from text prompts — Midjourney, DALL-E 3, Stable Diffusion 3.5, Flux, and Ideogram are the 2026 leaders.

AI technology that converts written text into natural-sounding spoken audio — the synthesis half of voice AI, distinct from STT which goes the other direction.

AI technology that generates video clips from text prompts — Runway Gen-4, OpenAI Sora, Google Veo, and Kling are the 2026 leaders. Output is typically 5–30 seconds.

The process of splitting text into subword units (tokens) that LLMs consume — a word like "tokenization" might become two or three tokens depending on the model's tokenizer.

The ability of an LLM to invoke external functions — APIs, shell commands, internal services — instead of just generating text.

A sampling strategy that picks from the smallest set of tokens whose cumulative probability exceeds p. Trims the long tail without hard top-k cutoff.

The neural network architecture that powers every modern LLM — uses self-attention to process sequences in parallel, replacing the older RNN approach for language modeling.

A reasoning pattern where the LLM explores multiple solution paths in parallel as a tree, evaluates partial paths, and backtracks — outperforming linear chain-of-thought on hard problems.

A pricing model where the customer pays for what they actually use — typically tokens, tool calls, compute minutes, or active agent hours — with no minimum seat commitment.

A database optimized for storing and querying high-dimensional embedding vectors — enabling semantic search, RAG, and similarity-based retrieval at scale.

A dense numerical vector representation of text, image, or audio — produced by an embedding model and used to measure semantic similarity in high-dimensional space.

Search powered by vector similarity — finding the K nearest embedding vectors to a query vector, typically using approximate-nearest-neighbor algorithms like HNSW or IVF.

Google Cloud's managed agent platform inside Vertex AI — visual builder, code SDK, native Gemini models, and A2A-protocol-first orchestration.

A 2025–2026 term for coding by describing what you want in natural language and letting AI generate the code — popularized by tools like Lovable, Bolt, Cursor Composer, and Replit Agent.

An agent capability for understanding images, screenshots, and video — letting the model reason over visual content.

A high-throughput open-source LLM inference engine — pioneered PagedAttention to manage KV cache like virtual memory, dramatically improving GPU utilization for serving open models.

An agent capability for taking phone calls, holding spoken conversations, and triggering actions from voice input.

An AI agent that takes phone calls, holds spoken conversations in real time, and triggers actions from voice input — handling customer support, scheduling, and outbound calling.

AI technology that creates a synthetic voice indistinguishable from a target speaker — typically trained on 30 seconds to a few minutes of clean source audio.

A benchmark of realistic web-task scenarios (e-commerce, social, content management) where agents are scored on completing real multi-step user goals through a real browser.

Short-lived, task-scoped memory the agent uses to track the current goal, plan, and intermediate results — analogous to a human scratchpad during a single problem.

An internal predictive representation of the environment that an agent uses to simulate the outcomes of candidate actions before acting — central to 2026 frontier-agent research.

An LLM's ability to perform a task it has never been explicitly trained on or shown examples of — relying entirely on the model's pre-training and the prompt instructions.