The AI concept atlas

Agent Consistency Gap part of Evaluation

The consistency gap is calculated from repeated evaluation runs to show how often average success fails to translate into repeatable success.

Agent contrasts with Workflow

This distinction concerns who directs the process: model decisions or primarily prescribed code.

Agent Harness contrasts with Evaluation Harness

An agent harness runs working interactions; an evaluation harness organizes tests and scoring.

Agent Harness often confused with Agent Runtime

Products bundle the loop and the durable execution layer under one offer, and vendors use overlapping names for them.

Agent Runtime enables Agent

Production agents rely on a runtime that keeps sessions alive and resumes work across failures and restarts.

Agent Runtime enables Memory

Durable state and checkpointing preserve agent state across restarts, which long-term memory builds on.

Agent Runtime enables Sub-agent

Runtime concurrency and queuing let several subagents execute in parallel under isolation.

Agent Runtime uses Orchestration

Some runtimes include orchestration features such as scheduling runs and coordinating multi-step workflows.

Agent Runtime uses Sandbox

Runtimes provide or integrate sandboxed environments where agent code and tools run under isolation.

Agent Runtime uses Trace

Runtimes record execution telemetry such as traces and audit logs so runs can be reviewed after the fact.

Agent Skill uses Tool

A skill can instruct an agent in using available tools; the tools still execute operations.

Benchmark part of Evaluation

A benchmark can supply a defined task set for an evaluation.

Computer Use uses Multimodal Model

The cited computer-use pattern interprets screenshots alongside text instructions.

Computer Use uses Tool Calling

A screenshot-based computer-use loop requests interface actions through tools.

Context Engineering uses Context Window

Context engineering selects information within the available context capacity.

Context Engineering uses Memory

Context management can select retained notes for a later model interaction.

Context Engineering uses Retrieval

Selecting external material through retrieval is one way to prepare useful context.

Context Window contrasts with Memory

Retained information is distinct from the information supplied in the current context.

Context Window uses Token

Context capacity is measured in tokens.

Delegation contrasts with Handoff

Delegation can return a subtask result; a handoff transfers control to another agent or stage.

Delegation part of Orchestration

Assigning worker tasks is one responsibility within multi-agent coordination.

Delegation uses Sub-agent

A coordinator can assign a bounded task to a sub-agent.

Evaluation contrasts with Reward

Evaluation measures behavior against criteria; a training reward supplies an optimization signal.

Evaluation Harness enables Evaluation

The harness runs tasks and collects evaluation results.

Evaluation Harness uses Benchmark

A harness can run benchmark tasks using shared scoring conventions.

Evaluation Harness uses Inference

Testing a language model requires running it on the evaluation inputs.

Evaluation uses Trace

Execution records can help inspect an agent evaluation trial.

Fine-tuning contrasts with Inference

Fine-tuning updates trainable parameters; inference runs the model on input.

Fine-tuning contrasts with RAG

Fine-tuning adapts parameters; RAG supplies retrieved evidence for generation. They can be combined.

Fine-tuning uses Language Model

Language-model fine-tuning starts with an existing model and continues training.

Guardrail part of Agent

An agent application can include checks around its inputs, outputs, or actions.

Handoff part of Orchestration

Transferring control between specialists is one coordination pattern.

Human-in-the-loop part of Agent

An agent workflow can pause selected actions for a human decision.

Human-in-the-loop part of Workflow

A workflow can pause a selected step until a person approves or edits it.

Inference uses Language Model

Inference runs a trained language model on supplied input.

Least Privilege part of Sandbox

Restricting a sandbox to necessary access applies least privilege to its permissions.

Least Privilege part of Tool

Tool permission design can grant only the operations needed for an assigned task.

LLM-as-a-Judge part of Evaluation

A model judge can supply judgments within a broader evaluation.

LLM-as-a-Judge uses Language Model

The judge is a language model prompted to assess outputs against criteria.

Memory uses Retrieval

An application can retrieve saved information when it needs to recall it.

Mixture of Experts contrasts with Sub-agent

An MoE expert is an internal network; a sub-agent is an application-level worker.

Mixture of Experts uses Token

A sparse language-model MoE routes tokens to selected internal experts.

Model Context Protocol enables Tool Calling

MCP can provide the integration through which an application discovers and calls tools.

Model Context Protocol enables Tool

MCP servers can expose tools to compatible AI applications.

Multimodal Model uses Language Model

In the cited LLaVA design, a vision encoder connects to a language model.

Orchestration uses Sub-agent

A manager-style orchestration can route tasks to specialist sub-agents.

Planning enables Delegation

A coordinator can use a plan to identify bounded tasks for workers.

Planning part of Agent

An agent may plan intended steps and revise that plan during its run.

Prompt Injection uses Retrieval

Retrieved documents can carry indirect instructions that attempt to redirect a model.

Prompt Injection uses Tool

An indirect injection may use a tool’s returned content as its delivery channel.

RAG uses Language Model

A generative model uses retrieved material to inform its output.

RAG uses Retrieval

Retrieval supplies external evidence to the generation process.

Reinforcement Learning uses Reward

The learning objective is expressed through rewards received from interaction.

Reinforcement Learning uses Rollout

Sampled trajectories provide interaction experience for reinforcement learning.

Agent uses Language Model

The model helps the agent choose its next action.

Agent Harness part of Agent

In this example, the harness runs the agent’s model and tool interactions.

Agent Harness uses Language Model

The harness invokes the model during execution.

Agent Harness uses Tool

The harness dispatches requested tool operations.

Agent Loop part of Agent Harness

The repeating interaction is implemented by the harness in this scoped example.

Agent Harness often confused with Scaffold

Some authors distinguish execution from configuration; others use overlapping boundaries.

Tool Calling uses Tool

A structured request identifies an exposed tool operation.

Agent uses Tool Calling

This tool-using agent requests actions and observes their results.

Agent Skill often confused with Tool

A skill packages task guidance and resources; a tool exposes an operation.

Rollout uses Inference

For a language-model policy, sampling a rollout includes model inference steps.

Sandbox enables Agent

An execution boundary lets an agent run tools within restricted access.

Sub-agent is a Agent

A sub-agent is an agent operating as a worker within a larger system.

Supervised Fine-tuning is a Fine-tuning

Supervised fine-tuning is fine-tuning with training targets such as desired responses.

Workflow uses Language Model

A prescribed workflow can use a language model for individual steps.