AI Agent News Glossary
This glossary explains the AI systems, products and measures that readers encounter on AI Agent News. An AI model can generate an answer or score a match. An AI agent can also choose actions to pursue a goal. A product can combine models, agents, fixed workflows and human approvals.
AI and models
- AI system
- A machine-based system that uses inputs to infer outputs, such as predictions, content, recommendations or decisions, in pursuit of objectives. Its autonomy and ability to adapt can vary. OECD AI Recommendation
- AI model
- A computational representation used to perform tasks such as classifying text, ranking results or generating a response. A model can be one component of an AI system.
- Machine learning (ML)
- Methods for fitting models to data or experience so they can perform tasks such as prediction or classification.
- Generative AI (GenAI)
- AI that generates content such as text, images, audio, video or code. NIST Generative AI Profile
- Foundation model
- A model trained on broad data that can be adapted to a range of downstream tasks. NIST glossary
- Large language model (LLM)
- A large model trained to process and generate language. An LLM can power an agent, a chatbot or a step in a workflow.
- Inference
- Using a model to produce outputs from inputs. Training and fine-tuning change model parameters.
- Prompt and context
- A prompt asks or instructs the model to do something. Context is the information available to the model for a call, such as instructions, messages, retrieved material or tool results.
- Embedding
- A numerical vector representing an item, such as a passage of text, for tasks like similarity search. MiniLM on Hugging Face is an example of a text embedding model. MiniLM model card
- Retrieval-augmented generation (RAG)
- A system retrieves relevant material from a separate store and supplies it to a generative model as context for a response. NIST glossary
- Grounding
- Tying an output to supplied information or observed facts. A citation identifies a source the reader can check for support.
- Multimodal AI
- An AI model or system that handles more than one type of input or output, such as text and images. State which modes the product supports.
- Model card
- Documentation accompanying a model, commonly covering intended use, evaluation, limitations and licence information. Hugging Face model cards
Agents and automation
- AI agent
- An AI system that pursues a goal, chooses at least some actions based on the situation and can carry them out within its permissions. An agent can pause for human approval. This is the site's working definition, based on the OECD's synthesis of goals, actions and autonomy. OECD agentic AI paper, section 3.2
- AI assistant
- A user-facing system that helps with tasks such as answering questions, drafting or taking approved actions. An assistant can include an agent.
- Agentic AI
- An umbrella term for systems that give one or more AI agents discretion over how they pursue goals. This is editorial usage. The OECD paper uses the term more narrowly for systems where multiple agents coordinate over longer periods. Specify the actions, supervision and agent count where known. OECD agentic AI paper, section 3.3 · Anthropic engineering guide
- Workflow and workflow automation
- A workflow is a set of steps for completing work. Workflow automation uses software to perform some or all of those steps. Describe an automated workflow as AI-assisted when AI is used within a predefined process, and describe it as agentic when an agent can choose how to carry out part of the work. Anthropic engineering guide
- Autonomy
- How much discretion an agent has over actions for a task. State what it may decide or do, where it may act and when it needs approval.
- Unattended execution
- A run that proceeds without a person operating it live. The run can use a fixed procedure, remain within pre-approved limits and stop or escalate on exceptions.
- Human oversight and approval
- Oversight is the ability to monitor and intervene. Approval is a decision that authorises a specific action or scope of work. Human-in-the-loop means a person participates in a decision or review step within the workflow. OECD explanatory memorandum
- Agent framework, harness and runtime
- A framework provides building blocks for developing an agent. A harness surrounds the model and manages instructions, tools, state and execution controls. A runtime provides the environment in which the agent runs. Products can provide several of these layers.
- Agent orchestration
- Coordination of steps, models, tools or agents, including routing work and handling dependencies. Multi-agent system specifies that more than one agent participates. Multiple model calls can be part of one agent or a fixed workflow.
- Agent memory and persistence
- Memory is information retained for later use. Persistent memory survives a session or run. A background agent can run while the user is elsewhere. A system may support persistence by staying active, waking for an event or resuming saved work. State which behaviour is supported.
- Tool and tool calling
- A tool is a callable capability, such as searching a database or updating a record. Tool calling lets a model or agent request that capability. The application controls execution and permissions.
- Model Context Protocol (MCP)
- An open protocol for connecting AI applications to capabilities such as tools, resources and prompts provided by servers. MCP specification
- Agent2Agent Protocol (A2A)
- An open protocol for communication and task exchange between independent agent systems. A2A specification
- Agent skill and Agent Skills
- An agent skill is a reusable package of task instructions and optional supporting material. Agent Skills is the named format that uses a
SKILL.mdfile. Use the product's own term for proprietary formats. Agent Skills specification - Computer use and desktop automation
- Computer use means model-directed interaction with computer interfaces, such as finding a control and clicking it. Desktop Automation is the site's broader Windows comparison, covering approaches such as scripts, replay, robotic process automation (RPA) and AI-guided tools. Desktop Automation
- Task mining and replay
- Task mining analyses recorded user activity to identify or examine tasks and recurring patterns. State whether capture is continuous or intentionally started. Replay runs a recorded or defined procedure.
- Personal agent and work agent
- Site labels for intended use. A personal agent acts on delegated tasks for an individual. A work agent is primarily marketed for workplace tasks. Products can serve both purposes. Personal Agents
Deployment, licences and evidence
- Deployment, managed and self-hosted
- Deployment describes where components run and who operates them. A managed service is operated by a provider. Self-hosted means the operator runs the specified component on infrastructure they control, which can include rented infrastructure. State where the model, agent runtime, data and required services operate.
- On-device and local inference
- On-device inference runs on the device itself. Local inference runs within the operator's local or self-hosted environment and can use a separate machine. State the location when it affects privacy or offline use.
- Fully self-hosted under the ADA benchmark
- The comparison's requirement that data, AI and execution operate under the user's control across the workflow. Check every required component, including any model services used during learning or repair. ADA Shortlist
- Open-source software
- Software distributed with source code and terms that meet the OSI Open Source Definition. Name the licensed component and its licence. Open Source Definition
- Open-weight model and Open Source AI
- Open-weight means the trained model parameters are available to obtain. Identify any access or use restrictions. Open Source AI, under the OSI definition, also requires freedoms to use, study, modify and share the system and access to the preferred form for modifying it, including relevant code, parameters and sufficient data information. Open Source AI Definition 1.0
- Evaluation (eval), benchmark and verification
- An eval checks behaviour on defined tasks. A benchmark sets evaluation criteria and can include test cases and scoring. Verification checks that a required outcome was achieved using relevant evidence. The ADA benchmark is AI Agent News's seven-criterion framework for desktop delegation. ADA Shortlist
- Evidence, availability and last checked
- Evidence is the material that supports a specific claim. A provider statement, demo and independently tested result support different levels of confidence. Availability tells readers who can access the product and under what release conditions. Last checked dates the review and should identify a product test only when one occurred.
- Microtask
- In the ADA benchmark, a small, reusable unit of desktop work that can be corrected, approved, verified and combined with other tasks. ADA Shortlist
Site views and measures
- Repository and project
- A repository is a versioned collection of code or other files. A project can span repositories, models or products. Hugging Face uses repositories for models, datasets and Spaces. A Space can host an app or demo. Hugging Face repositories
- GitHub Trends
- Saved GitHub projects ordered by observed star growth per day, comparing saved observations roughly a week apart. The dates on each card give the actual comparison period. GitHub Trends
- GitHub Popular
- Saved GitHub projects ordered by their latest saved star count. A star indicates that a GitHub user chose to star the repository. GitHub Popular · GitHub stars
- Hugging Face Trends
- A saved snapshot of Hugging Face's trending model ranking. The cards also show platform metrics such as likes and downloads. Hugging Face Trends
- Hugging Face Popular
- A saved ranking of models by platform-reported downloads over the 30 days before the fetch. Counts can include repeat and automated requests. Hugging Face Popular · Hugging Face download counting
- YouTube Trends, archive and analysed sample
- YouTube Trends describes discussion in the site's saved video and transcript collection. The archive counts stored videos. The analysed sample counts records that met the requirements for the particular chart or review and were included in it. Use the denominator, date range and coverage defined for that analysis. YouTube Trends
- Attention, adoption and quality
- Attention describes measurable interest or activity on a specified platform. Adoption requires evidence of actual use by people or organisations in a defined setting. Quality depends on outcomes for the task or requirements being assessed.
- Editorial card and cross-reference
- An editorial card is an AI Agent News summary or assessment of an item. A cross-reference links an item to related evidence, such as a GitHub implementation of a model or YouTube coverage of a product. State the relationship the evidence supports.
Terminology sources
- International standard. ISO/IEC 22989:2022 is the baseline for general AI terminology. The review used the public ISO listing and public terminology references. ISO/IEC 22989 · OIML terminology
- Policy and technical references. The OECD AI Recommendation provides the AI system definition. The OECD agentic AI paper and Anthropic's engineering guide illustrate different uses of agentic AI. NIST provides technical definitions and risk guidance. OECD AI Recommendation · OECD agentic AI paper, sections 3.2–3.3 · Anthropic engineering guide · NIST Generative AI Profile
- Specifications and licensing. Use the published specifications for A2A, MCP and Agent Skills, and the Open Source Initiative (OSI) definitions for claims about open-source software or AI. A2A · MCP · Agent Skills · Open Source Definition · Open Source AI Definition 1.0
- Platform definitions. Use GitHub and Hugging Face documentation for their repository types and metrics, and AI Agent News methodology for how saved observations are presented. GitHub stars · Hugging Face repositories · Hugging Face model downloads