The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →An agentic AI system is more than a language model: it combines a model with instructions, orchestration, tools, permissions and an environment where it can take actions. Some systems follow fixed workflows; others let a model decide what to do next. This glossary explains the terms behind those designs and where their meanings can vary by provider.
What is an AI agent?
AI agent
In this glossary, an AI agent is a system that uses a model to pursue a goal by processing input, deciding what to do, and sometimes using tools to take actions. Google Cloud describes agents in terms of goals, reasoning, available tools and actions. Anthropic emphasizes a model directing its own process and tool use. These are useful descriptions, not a single definition adopted by every vendor.
As an Amazon Associate I earn from qualifying purchases.
Agentic AI
Agentic AI describes systems designed to take steps toward goals rather than only produce a one-off response. The term covers a range: a tightly constrained system may execute a predefined sequence, while a more dynamic one may choose among steps and tools as it works. The label alone does not tell you how much autonomy a product has.
AI agent vs. chatbot
A chatbot typically responds to a user’s prompt in conversation. An agent may also plan a sequence of actions, call tools, use returned information and continue until it reaches a stopping point. The distinction is about system behavior, not the chat interface: an agent can appear as a chatbot, and a chatbot can have limited tool access without being highly autonomous.
How is an agent different from a workflow?
Workflow
A workflow follows steps and branches defined in advance. Anthropic summarizes the distinction this way: “Workflows are systems where LLMs and tools are orchestrated through predefined code paths.” This quote is from its engineering guidance, Building Effective AI Agents, published December 19, 2024.
Agent
In Anthropic’s framing, an agent dynamically directs its process and tool use instead of following only a fixed path. It can decide which step to take next based on the task and what it learns along the way. Dynamic choice can support flexible work, but it also makes behavior less predictable than a fully predefined sequence.
Workflow and agent continuum
These are architectural patterns, not always mutually exclusive product categories. A system can use a fixed workflow for routine steps and let a model choose among tools or routes at selected points. To understand a particular system, ask which decisions are code-defined, which are model-directed, and where a person can review or stop an action.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Fix the driver behind crashes, sound loss and screen glitches3Repair Windows errors before they cause bigger problemsWhat do model, LLM and context window mean?
Model
A model processes inputs and produces outputs. In an agent system, the model is one component: its behavior also depends on instructions, orchestration, available tools, permissions and the environment it operates in.
Foundation model
A foundation model is a broadly trained model that can be adapted to different tasks. Google Cloud distinguishes foundation models by the types of material they can handle, including text, images, audio and video.
Large language model (LLM)
An LLM is a foundation model focused on language and text-based tasks. It may serve as the reasoning or language component in an agent, but it is not, by itself, the whole agent system.
Token
A token is a unit used to represent text for a model. Tokens are not always whole words: depending on how text is divided, a word may be one token or several.
Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →Context window
A context window is the amount of input a model can process at a given time, measured in tokens. It can include instructions, conversation history, retrieved material and tool results. It is a limit on current processing, not a guarantee that the system will retain information between sessions.
Memory
Memory is information kept for later use, such as a saved preference or a record of prior interactions. Short-term context is what the model can process in the current request; persisted memory is stored separately and may be selected or retrieved for later requests. A system may have a context window without persistent memory, or memory that is much smaller than its context window.
How does an agent plan and use tools?
Orchestration
Orchestration is the control layer that coordinates the model with planning, state, memory, tools and data flow. It determines how requests and results move through the system and can enforce rules about which actions are allowed.
Agent loop
An agent loop is the repeated cycle of assessing the task, choosing a step, using a tool or producing an answer, and evaluating what happened. A system may continue the loop when a tool result indicates more work is needed, or stop when it reaches a defined completion or safety condition. The loop’s exact design varies by implementation.
Tool
A tool is an external capability an agent can call, such as searching a data source or invoking a service. A tool’s permissions define what the agent can actually do; having a tool available does not give the model unlimited access to other systems.
Function calling
Function calling is a model-mediated request to run a named function with specified inputs. The model proposes or selects the call; the surrounding system validates and executes it, then can return the result to the model. It is a mechanism for connecting a model to capabilities, not evidence that the model can execute arbitrary code.
State
State is information about the system’s current task or execution, such as completed steps, pending actions or tool results. Orchestration can use state to decide what happens next; it is distinct from a model’s weights and from any long-term memory a system stores.
What do RAG and grounding mean?
Retrieval-augmented generation (RAG)
RAG means retrieving relevant material, adding it to the model’s context, and generating an answer using that material. It can help a system work with current or specialized information that is not contained in the model’s built-in knowledge. The answer still depends on whether the system retrieves the right material and whether that material is accurate and relevant.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Rank #4
Retrieval quality and source quality
These are separate checks. Retrieval quality asks whether the system found useful evidence for the question. Source quality asks whether that evidence is trustworthy, current and appropriate. A well-retrieved passage can still be wrong or outdated, and a weak retrieval can leave a model answering without the evidence it needs.
Grounding
Grounding connects a model’s output to data or evidence, such as retrieved documents or tool results. It can make an answer easier to verify, but it does not certify that the evidence is correct or eliminate unsupported statements. Check whether the cited material actually supports the specific claim being made.
What are a harness and prompt injection?
Harness
Anthropic uses “harness” for the instructions and guardrails around a model in an agent setup. The harness can shape how the model behaves and how the system handles tool use, but the protections also depend on implementation, permissions and the environment.
Prompt injection
Prompt injection is an attempt to redirect a model by placing malicious instructions in material it is asked to process, such as retrieved content. The injected text may try to override the task or induce an unsafe action. Treat untrusted content as data rather than authority, restrict what tools can do, and use layered defenses and review for consequential actions. No single safeguard guarantees protection.
Permissions
Permissions determine which resources and actions an agent can access. Limiting access to what a task requires reduces the potential impact of mistakes or malicious instructions. A useful review asks not only what the model is instructed to do, but also what its connected tools are technically allowed to do.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Where do people and protocols fit?
Human-in-the-loop (HITL)
Human-in-the-loop describes a design where a person reviews, guides or makes decisions during an automated process. The phrase is broad; it does not specify whether the person sees every step or only selected decisions.
Approval checkpoint
An approval checkpoint is a specific pause before an action that requires authorization, such as sending a message or changing a record. A useful checkpoint tells the reviewer what the agent intends to do and gives them a way to approve, reject or revise it. It is especially important when an action is consequential or difficult to reverse.
Model Context Protocol (MCP)
MCP is an open standard for connecting models to external data sources and tools, according to Anthropic. Anthropic has said it donated MCP to the Linux Foundation’s Agentic AI Foundation. MCP describes a way to make connections; it does not by itself establish what a particular implementation can access, whether a server is trustworthy, or what safeguards are in place.
Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minutePC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11How are agent systems evaluated and debugged?
Evaluation
Evaluation measures how a system performs on tasks and how it behaves while doing them. Useful dimensions include task completion, reliability across repeated runs, tool-use quality, safety, latency, operating cost, context limits, integration fit and the amount of human oversight required. A single successful run does not establish reliable performance across different inputs.
Tool-use quality
Tool-use quality concerns whether the system chooses an appropriate tool, supplies suitable inputs, interprets the result correctly and avoids unauthorized actions. It is distinct from whether the final answer sounds plausible.
Trace
A trace records steps and interactions during an execution, such as model calls and tool requests. It can help developers investigate where a run went wrong or why it took a particular route. A trace is evidence about that execution, not proof that the system will behave the same way on every task.
Safety evaluation
Safety evaluation examines whether a system handles risks appropriately, including unsafe outputs or tool actions. Google Cloud’s agent-evaluation documentation lists response quality, tool-use quality, hallucination and safety metrics, and marks the evaluation feature Preview. That status means availability should not be assumed for every user, region or deployment.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
