Capabilities terms

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.

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.

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.

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.

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.

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.

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.

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.

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.

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.

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

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 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.

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

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 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.

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.

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

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

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.

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.

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

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.

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.

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.

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).

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

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.

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

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.

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.

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.

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.