iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
An AI agent is not just a language model: it is a system that combines a model with instructions, orchestration, tools, permissions, and an environment in which it can act. This practical glossary explains the terms that matter when you build, evaluate, or use agentic AI—and distinguishes concepts that vendors sometimes describe differently.
What is an AI agent?
In this article, an AI agent is a system that uses a model to work toward a goal by processing input, deciding what to do, and using available tools or taking actions. Google Cloud describes agents in terms of goal-directed reasoning, tools, and actions. Anthropic emphasizes a model directing its own process and tool use. These are useful descriptions, not a single definition that every provider follows.
An agent’s behavior depends on more than its model. Instructions, orchestration, tools, permissions, and the environment all shape what it can do. An LLM may be the model component, but it is not the complete agent system.
Free tools Windows power users keep installed
One-click scans. No signup required.
How does an AI agent differ from a chatbot or a workflow?
A chatbot is an interface for conversation; it may simply respond with text, or it may be connected to tools and operate as part of an agent system. The label alone does not tell you how much the system can do.
Anthropic draws an architectural distinction: “Workflows are systems where LLMs and tools are orchestrated through predefined code paths.” In contrast, an agent can dynamically direct its process and select tools. Real systems often combine both patterns, so workflow and agent are better understood as design choices than as mutually exclusive product categories.
| Design | How steps are selected | Typical trade-off |
|---|---|---|
| Workflow | Predefined code paths determine the sequence. | More predictable control over the process, but less flexibility when a task needs an unanticipated path. |
| Agent | The model can dynamically direct steps and tool use. | More adaptable for multistep tasks, but behavior and tool choices need careful evaluation and control. |
| Hybrid | Code fixes some steps while the model selects among others. | Can balance structure and flexibility; the system’s actual boundaries depend on its implementation. |
What do model and prompt terms mean?
Foundation model
A foundation model is a broadly capable model that can be adapted or used for different tasks. Google Cloud distinguishes models focused on text from models that can work across modalities such as text, images, audio, and video.
Large language model (LLM)
An LLM is a text-based foundation model. In an agent system, it may interpret a request, decide on a next step, or produce a response, but the surrounding system controls what tools and actions are available.
Prompt and instructions
A prompt is input supplied to a model, which can include the user’s request and other context. Instructions tell the system how it should behave or what constraints to follow. Instructions guide behavior; they do not by themselves grant access to data or tools.
Harness
Anthropic uses harness for the instructions and guardrails surrounding a model in an agent setup. A harness is part of the system around the model, not another name for the model itself.
What are orchestration, an agent loop, and a trace?
Orchestration
Orchestration is the control layer that coordinates planning, state, memory, tool use, and data flow. It determines how the system moves between model calls and other parts of the application.
Agent loop
An agent loop is the repeated process of considering the current task, choosing a next step, and using the result to decide what to do next. A system may continue until it reaches a stopping condition, completes the task, or needs a person to intervene. The loop’s exact behavior depends on the design; the term does not imply unlimited or unsupervised action.
State
State is information about the task or execution that the system carries forward as it proceeds. It may include what has happened so far or what remains to be done. State is managed by the application and should not be confused with a model’s built-in knowledge.
Trace
A trace records steps and interactions during an execution, such as model activity and tool use. It can help diagnose where a run went wrong, but one trace is evidence about that run—not proof that the system will perform reliably in other cases.
What are tools, function calling, and permissions?
Tool
A tool is an external function, API, service, or other capability an agent can use. Depending on its configuration, it might retrieve information or perform an action. The tool—not the model alone—carries out the external operation.
Function calling
Function calling is a way for a model to request that an application run a specified function. The application handles the request and may return the result to the model. It is not unlimited access: the functions exposed to the model and the application’s checks determine what can be called.
The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →API
An application programming interface (API) is a defined way for software to request data or actions from another service. An agent can use an API when its application makes that connection available.
Permissions
Permissions determine which tools, data, and actions the system is allowed to access. Limiting permissions to what a task requires reduces the consequences of an incorrect decision or a malicious instruction; a model’s apparent ability to use a tool is not a substitute for access control.
What is a context window, and how is it different from memory?
Token
A token is a unit of text processing used by a model. It may represent a word, part of a word, or other text content; it is not necessarily the same as a word or character.
Context window
A context window is the amount of tokenized input a model can process for a given interaction. It may contain the current prompt, conversation content, instructions, and retrieved material. It describes current processing capacity, not durable storage.
Rank #4
Short-term context
Short-term context is information available to the model within its current interaction or task, typically through the content provided in the context window. As that content changes or a new interaction begins, what is available may change too.
Persisted memory
Persisted memory is information an application stores and can make available again later. Whether an agent has it, what it retains, and how it is updated are implementation choices; a large context window does not automatically create persistent memory.
What do RAG and grounding mean?
Retrieval-augmented generation (RAG)
RAG combines information retrieval with generation: the system retrieves relevant material, adds it to the model’s context, and then generates a response. It can help an application draw on specialized or current material rather than relying only on what the model learned during training.
Retrieval quality and source quality are separate concerns. A retrieval system can surface irrelevant or incomplete material, and relevant material can itself be inaccurate or outdated. Supplying sources to a model does not certify the resulting answer.
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 problemsRetrieval
Retrieval is the step that finds and returns information for a query. The usefulness of RAG depends in part on whether the retrieval step finds suitable material for the task.
Grounding
Grounding connects a model’s output to data or evidence supplied to the system. It can make the basis for an answer more inspectable, but it does not guarantee that the evidence is correct or that the model interpreted it accurately.
Hallucination
A hallucination is an inaccurate or unsupported model output presented as if it were reliable. Retrieved material or other grounding may help a system answer from evidence, but neither removes the need to check the sources and the answer.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What is prompt injection, and why do approvals matter?
Prompt injection
Prompt injection is an attempt to redirect a model by embedding malicious instructions in material it is asked to process. For example, content retrieved for a task could contain instructions that conflict with the user’s intent. An agent’s access to tools and data affects the possible impact, so security depends on layered defenses and tightly scoped permissions—not on one safeguard alone.
PC 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 & 11Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchHuman-in-the-loop
Human-in-the-loop describes a system design in which a person reviews, guides, or intervenes in the process. This can be useful where context or judgment is important, but the phrase alone does not specify when review happens or what a reviewer can control.
Approval checkpoint
An approval checkpoint pauses a process for human review or authorization before a consequential action. It is a specific control point; it should be placed where the consequences of an error justify the interruption.
What is MCP?
The Model Context Protocol (MCP) is an open standard Anthropic describes for connecting models to external data sources and tools. An implementation’s actual capabilities depend on the servers and connections it exposes, as well as the permissions applied to them. Protocol support alone does not mean a model can access every external service.
How should you evaluate an agent?
Assess the system on the task it is meant to perform, not on the agent label or a single impressive run. A useful evaluation considers both the final result and the steps taken to reach it.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →- Task completion quality: Did the output meet the task’s requirements?
- Reliability: Does it perform acceptably across repeated runs and relevant variations?
- Tool-use quality: Did it choose appropriate tools and use them correctly?
- Safety and permissions: Did it stay within allowed data and actions, including when given misleading inputs?
- Latency and operating cost: Are the time and resources required suitable for the use case?
- Context limits and integration fit: Can it handle the required information and work with the surrounding systems?
- Human oversight: Are review and approval points appropriate to the consequences of errors?
Evaluation should include intermediate behavior, such as tool choices and safety-related decisions, as well as the final answer. Google Cloud’s evaluation documentation lists response quality, tool-use quality, hallucination, and safety among its metrics; its evaluation feature is marked Preview, so availability should not be assumed across Google Cloud products or users.
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.
