Find the term.
See how it fits.
Explore the concepts behind AI agents, one useful connection at a time.
Concept directory
Agent
An agent is a system that uses a model to choose actions toward a goal, then responds to what happens. How much freedom it has depends on the application.
Agent Harness
An agent harness runs interactions between a model and its tools. It supplies context, routes action requests, returns results, and manages when execution continues or stops.
Agent Loop
An agent loop repeats a cycle of model decisions, actions, and feedback until a stopping condition is met. It lets the system adapt its next step to the result of an earlier action.
Agent Skill
An Agent Skill packages instructions and optional resources for a recurring task. A compatible agent can discover it, load its instructions when needed, and use the included resources.
Benchmark
A benchmark is a defined set of tasks and scoring conventions used to compare performance. An AI benchmark measures selected capabilities under its particular test conditions.
Computer Use
Computer use lets an AI system interact with software through a graphical interface. A common implementation uses screenshots for observation and tool actions such as clicks, typing, and scrolling.
Context Engineering
Context engineering is the practice of selecting, organizing, and maintaining the information a model receives while it works. It includes instructions, tool descriptions, retrieved material, and conversation state.
Context Window
A context window is the token capacity available for a model’s current interaction. It limits how much supplied information and generated content can fit in that interaction, according to the model and service’s accounting rules.
Delegation
Delegation assigns a task or responsibility to another worker. In an agent system, a coordinating agent can give a specialist a bounded job and use the returned result.
Evaluation
Evaluation is checking an AI system against defined tasks and success criteria. An evaluation may score an answer, inspect actions, or test whether the resulting environment matches the intended outcome.
Evaluation Harness
An evaluation harness is software that runs evaluation tasks against models or systems and collects results. It connects task definitions, model execution, and scoring into a repeatable testing process.
Fine-tuning
Fine-tuning continues training from an existing model to adapt it to data, a task, or a domain. It changes trainable parameters, rather than only changing the input supplied for one request.
Guardrail
A guardrail is a check or control applied around an AI system’s inputs, outputs, or actions. It can block, transform, or route behavior according to a rule or a model-based assessment.
Handoff
A handoff transfers control from one agent or stage to another. In a conversational system, it can change which specialist is responsible for handling the next interaction.
Human-in-the-loop
Human-in-the-loop introduces a person’s decision into an automated process. In an agent workflow, execution can pause at a review point so a person can approve, change, or reject a proposed action.
Inference
Inference is using an already trained model to produce an output from an input. For a text-generating language model, this usually means processing a prompt and generating tokens in sequence.
Language Model
A language model learns patterns in language and uses supplied context to produce a response. In an agent application, it can propose the next action; surrounding software carries out that action.
Least Privilege
Least privilege means giving an entity only the access and authority it needs for its assigned task. For an agent application, this can limit the tools, data, or operations available to it.
LLM-as-a-Judge
LLM-as-a-judge uses a language model to assess outputs against instructions or a rubric. It can produce scores, comparisons, or written judgments to support evaluation.
Memory
In agent applications, memory is retained information that can be reused later. It can include conversation state within a session or information saved across sessions, depending on the application.
Mixture of Experts
A mixture-of-experts model uses multiple internal expert networks and a routing mechanism. In a sparse language-model implementation, the router sends each token through only a subset of those experts.
Model Context Protocol
Model Context Protocol is a standard for connecting AI applications to external capabilities and information. It defines how compatible clients and servers exchange access to features such as tools, resources, and prompts.
Multimodal Model
A multimodal model works with more than one kind of information, such as images and text. Which combinations it accepts or produces depend on the particular model.
Orchestration
Orchestration coordinates the order, routing, and results of work across components. In a multi-agent system, it can determine which agent receives a task and how returned results are combined.
Planning
Planning organizes intended steps toward a goal before or during execution. An agent may create a plan, use it to guide actions, and revise it as new information arrives.
Prompt Injection
Prompt injection is an attempt to redirect an AI system through instructions embedded in input or external material. The system may treat that material as authoritative instructions instead of data to process.
RAG
Retrieval-augmented generation combines a generative model with retrieved external information. The retrieved material helps inform the response instead of relying only on information encoded in the model’s parameters.
Reinforcement Learning
Reinforcement learning trains an agent’s behavior using reward signals from interaction. The objective concerns expected cumulative reward, so an action’s value can depend on consequences that occur later.
Retrieval
Retrieval is finding relevant information from an external collection in response to a query. Its output can be documents or passages that an application supplies to a model.
Reward
A reward is a feedback signal used to guide learning or assess outcomes against an objective. In language-model training, a learned reward model can assign scores based on preference examples.
Rollout
In reinforcement learning, a rollout is a sampled sequence of interaction: observations or states, actions, and resulting rewards. It provides experience for analysis or learning.
Sandbox
A sandbox is an execution environment that restricts what running code can access or change. For an agent’s tools, boundaries can include allowed files, directories, and network destinations.
Scaffold
Scaffold can describe the instructions, tool descriptions, and context setup that shape a model’s behavior. Some authors use the word for more of the surrounding agent software.
Sub-agent
A sub-agent is an agent assigned a bounded role within a larger agent system. In a supervisor pattern, it receives work from a coordinating agent and returns a result.
Supervised Fine-tuning
Supervised fine-tuning adapts a pretrained model using examples of desired behavior. For language models, training can teach the model to produce target responses or continue supplied text.
Token
A token is a unit that a tokenizer turns into an identifier for a model to process. In text models it may represent a word, part of a word, punctuation, or another text fragment.
Tool
A tool is an operation an AI application exposes for the model to request, such as searching documents or looking up a record. Its interface defines the information needed to use it.
Tool Calling
Tool calling lets a model request an operation through a structured interface. The application executes the request and makes the result available to the model.
Trace
A trace records the operations involved in an execution so their sequence and relationships can be inspected. In distributed tracing, individual operations are represented as spans with timing and other recorded attributes.
Workflow
In this Atlas, a workflow is an arrangement of steps whose control flow is primarily specified by application code. It may include language models, tools, branching, and repeated operations.