An AI agent uses a language model, context, and available tools to work toward a goal through one or more steps. The useful way to understand agentic AI is as a cycle: gather context, decide what to do, act, inspect the result, and continue or stop. Here are 20 terms that map that cycle and the design choices around it. This is a practical glossary, not a canonical list; “agentic” describes a degree of autonomy, not a single standard architecture.
Start with what an agent does
1. Agent
An agent is software that uses a language model and tools to pursue a goal. Microsoft Visual Studio Code puts it simply: “An agent is an AI system that uses a language model and tools to complete a goal on your behalf.” The model alone is not necessarily an agent: the surrounding application supplies context, exposes capabilities, runs requested actions, and handles their results. Microsoft Visual Studio Code’s agent concepts documentation
2. Agentic
“Agentic” describes a system or workflow with some ability to make decisions and act toward a goal. It is a matter of degree, not a yes-or-no product category. One system may choose among tools and revise its approach; another may follow a tightly constrained sequence while still using an agent-style model call.
3. Agentic workflow
An agentic workflow is a process in which an agent plans or takes actions toward a goal and can adjust its next step based on feedback. That distinguishes it from a fixed workflow, where the same predefined steps run in the same order. Real systems can mix the two: fixed steps may set boundaries while the agent chooses what to do within them.
#1 Best Overall
4. Agent loop
The agent loop is the repeated cycle of interpreting context, deciding, acting, and evaluating what happened. Google for Developers names its typical stages “Observe,” “Reason,” “Act,” and “Feedback.” In code, that might mean reading a user request, deciding to inspect a file, receiving the file contents, and then choosing whether to edit it or ask a question. Google’s Machine Learning Glossary: Agentic
5. Planning
Planning means selecting steps to reach a goal. A plan-and-solve approach drafts several steps before acting, but a plan is not a promise that every step will remain appropriate: new tool results may require the agent to revise its next move. For a constrained task, a short plan can improve clarity; for a changing task, the ability to adapt matters more than sticking rigidly to an initial outline.
Understand the capabilities an agent can use
6. Tool
A tool is a capability the agent can request, such as searching a knowledge base, reading a file, or calling an API. The surrounding application or runtime—not the language model by itself—executes the request and returns a result. Tool design therefore determines not just what the agent can do, but what data and actions are reachable.
7. Tool calling or function calling
Tool calling is the structured way a model requests a named capability and supplies its parameters. For example, it might request a weather lookup with a city field. The application validates and executes that request, then passes the result back into the conversation or workflow. Function calling is commonly used for the same pattern; the exact label depends on the platform. A request to call a tool is not itself proof that the action succeeded.
Recommended Free Tools
Rank #2
8. Action space
An agent’s action space is the set of tools, resources, and permissions available to it. Too many poorly bounded options can make selection more error-prone; too few may prevent task completion. Google’s guidance treats action-space design as a balance between capability and constraint. For developers, the practical question is not only “Which tools should be exposed?” but also “What can each tool read or change, and under whose authority?”
9. Autonomy
Autonomy is how much the system can plan, act, and adapt without a person intervening at every step. It is a spectrum shaped by both workflow and permissions. An agent that can decide which read-only search to run has less consequential freedom than one allowed to send messages, modify records, or spend money. Describe autonomy in terms of actual decisions and permissions, not as a blanket claim that an agent works independently.
Know how agent systems are coordinated
10. Orchestration
Orchestration coordinates and routes work among model calls, tools, agents, or workflow steps. It can be a fixed sequence written by a developer or a runtime decision about which component should handle the next task. Orchestration does not by itself mean several autonomous agents are involved; a single agent’s tool calls can also be orchestrated.
11. Subagent
A subagent is a narrower specialist agent assigned part of a larger task, often by a manager agent or orchestrator. A coding assistant, for example, could delegate test discovery to one specialist and documentation review to another. Delegation can divide work, but it also adds coordination: the system must provide useful context, combine results, and handle disagreements or gaps.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Rank #3
12. Multi-agent system
A multi-agent system uses multiple specialized agents that collaborate or pass work between them. It is one architectural option, not an automatic upgrade over a single agent with several tools. Multiple agents can separate responsibilities, but also create more handoffs and more places where context can be lost. AWS documents both single-agent and multi-agent patterns. AWS Agentic AI Lens definitions
13. MCP (Model Context Protocol)
MCP is an open protocol for standardizing connections between AI applications or agents and external tools, data, and services. It can support discovery of tools, prompts, and resources, with authorization controls described in Google Cloud’s documentation. MCP is a connection protocol; it is not the tool itself, a guarantee that a model will use a tool correctly, or a substitute for application-level permission checks. Google Cloud’s remote-server documentation reports support for protocol version 2026-07-28; version support is implementation-specific and can change. Google Cloud MCP servers overview
Separate retained information from retrieved information
14. Agent memory
Agent memory refers to mechanisms for retaining and retrieving information across steps or sessions. Session memory is temporary context associated with an ongoing interaction; persistent memory can carry information forward. AWS also describes memory by content or purpose, including episodic memory (past events), semantic memory (facts and concepts), and procedural memory (how to perform a task). Storing information does not guarantee that it will be relevant, current, or safe to use later. AWS Agentic AI Lens definitions
15. RAG (retrieval-augmented generation)
RAG supplies retrieved material as context for a model’s response. In a basic implementation, retrieval may happen in a fixed preprocessing step before the model generates an answer. A retrieval-augmented system can ground responses in an external knowledge source, but the quality of the result still depends on what was retrieved and how the model uses it.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Scan for outdated or missing drivers - takes under a minute3Clear out junk files and repair common Windows errors16. Agentic RAG
Agentic RAG puts retrieval decisions inside the agent’s reasoning loop. The agent can decide whether it needs more information, choose what to retrieve, select a retrieval tool, and assess whether the returned context is sufficient. The distinction is about control: RAG describes using retrieved context, while agentic RAG lets the agent make adaptive choices about retrieval. AWS Agentic AI Lens definitions
17. Embedding
An embedding is a numeric vector representation of text. Systems often use embeddings to find content that is semantically similar to a query, even when it does not share the same exact words. Embeddings can support semantic search in a RAG pipeline; they are a representation used for retrieval, not the retrieved answer or a guarantee of relevance. Mendix glossary
Set boundaries and check the work
18. Human in the loop
Human-in-the-loop design creates a defined point where a person can approve, correct, or decide before work proceeds. This is especially important for consequential or hard-to-reverse actions. A useful approval gate is specific about what the person is authorizing; a vague “review the agent” step can be difficult to apply consistently.
19. Evaluator or critic
An evaluator, sometimes called a critic, checks an output before it is finalized. It may look for missing requirements, unsupported claims, or a failed test. Evaluation can catch problems, but it does not guarantee correctness: the evaluator can miss an issue or make the same mistaken assumption as the agent. Pair checks with appropriate tests, evidence, and human review for high-impact work.
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 matchBest Value
20. Termination condition
A termination condition is a defined rule for ending the loop. Examples include successful completion, exhausted time or tool budget, or a person identifying a problem. Without a stopping rule, an agent may continue taking unnecessary steps; with an overly strict one, it may stop before completing the task. Google’s glossary includes termination conditions among the core agentic concepts. Google’s Machine Learning Glossary: Agentic
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How the concepts fit together in practice
Suppose a developer asks an assistant to find why a test is failing and propose a fix. The agent interprets the goal and plans an initial investigation. Its action space includes read-only file access and a test runner. It uses tool calling to inspect the error and run tests; the runtime returns the results. The agent loop uses that feedback to decide whether to inspect another file, revise its plan, or stop. A fixed workflow might prescribe the same diagnostic steps each time; a more adaptive agent can choose among them based on what it observes.
If the assistant needs project documentation, RAG can provide relevant passages. With agentic RAG, it may decide which documentation query to run and whether the results answer the question. Memory may preserve a session detail or a durable preference, while embeddings can help locate semantically related documentation. MCP may provide a standardized way to connect the application to external capabilities, but the application still needs to control permissions. If the assistant is allowed to edit files, a human approval step can gate that action; an evaluator can check the proposed change, and a termination condition can stop the work once the requested result is reached.
Choose the architecture to match the task
| Design choice | Useful when | Main trade-off |
|---|---|---|
| Fixed workflow or state machine | The sequence is predictable and the allowed paths should be tightly controlled. | Generally more constrained and less adaptable outside its predefined rules. |
| Adaptive agent behavior | The next useful step depends on context or tool results. | Requires careful limits, evaluation, and stopping conditions because the path can vary. |
| One agent with multiple tools | A single decision-maker can handle the task using distinct capabilities. | Responsibilities remain centralized; tool selection and permissions still need design. |
| Multiple agents with orchestration | The work benefits from distinct specialist responsibilities or separable tasks. | Introduces handoffs and coordination overhead; more agents do not automatically improve outcomes. |
| Temporary session context | Information is needed only during the current interaction. | It may not be available in a later session. |
| Persistent memory | Relevant information should carry across sessions. | Stored information can become stale or inappropriate; retention and retrieval need controls. |
Google notes that constrained state-machine agents generally make fewer mistakes but adapt less freely outside their rules. AWS describes single-agent and multi-agent system patterns. Neither comparison establishes a universally superior design: choose based on task variability, impact of mistakes, permissions, and how much human oversight is appropriate. Google’s Machine Learning Glossary: Agentic · AWS Agentic AI Lens definitions
Free tools Windows power users keep installed
One-click scans. No signup required.
Keep the vocabulary straight
- Tool is the capability; tool calling is the structured request to invoke it; MCP is one protocol for connecting applications to external tools or data.
- Memory retains information; RAG retrieves information to provide context. Agentic RAG makes retrieval itself adaptive.
- Orchestration coordinates steps or components; it does not necessarily imply multiple agents.
- Agentic describes a degree of autonomy; it does not specify one universal architecture.
For a narrower, MCP-focused reference, O’Reilly’s The MCP Standard: A Developer’s Guide to Building Universal AI Tools with the Model Context Protocol includes a glossary appendix. O’Reilly Media, The MCP Standard
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.
