AI Agent Glossary: 25 Terms Every Team Should Know

Teams building AI agents run into the same problem early: people use the same words to mean different things. One person's "memory" is another's "knowledge base". "Approval" might mean a button in the UI or a hard stop in the runtime. Those mismatches turn into wrong estimates, confused reviews and features nobody asked for.

This glossary defines 25 terms that come up in almost every agent project, grouped into four areas: agents and runtime, knowledge and memory, control and safety, and models and deployment. Each definition is short and practical, and where it helps, notes how the term maps to LibraOS.

Agents and runtime

1. AI agent. Software that uses a language model to pursue a goal: it decides which steps to take, calls tools, uses context, and keeps going until the task is done or it needs a person. A chatbot replies; an agent completes tasks.

2. Agent framework. A library you write code against to build agents: prompts, tools, chains and coordination. The agent runs inside your own application, and you build what surrounds it.

3. Agent runtime. A running layer your application calls that executes agents and holds what they depend on. LibraOS describes itself as the AI runtime for your apps: the execution layer that runs agents, memory, knowledge and tools beneath your own interface.

4. Tool call. A request from the agent to a function or system outside the model, such as looking up an order or sending a message. A tool call that sounds made in the reply but never reached the system is one of the quietest agent failures.

5. Side effect. Any action that changes something outside the conversation: updating a record, spending money, sending a message. Side effects are what approval points exist for.

6. Structured output. A reply that follows a defined schema, such as JSON with fixed fields, so your code can act on it without parsing free text.

7. Long-running task. Work that outlives a single request, such as reviewing hundreds of records. It needs saved progress, safe retries, clear status and a way to pause for people. Recovery after an interruption depends on the workflow; don't assume every run resumes on its own.

Knowledge and memory

8. Context window. The amount of text a model can consider at once, measured in tokens. Anything outside it, such as older messages in a long conversation, isn't seen by the model unless something brings it back in.

9. Conversation history. The earlier messages of the current conversation, sent along so the model can follow it.

10. Agent memory. Context carried forward from earlier interactions, so the agent doesn't start from zero: facts, decisions, preferences or task state from previous sessions. In LibraOS, memory carries useful context across sessions.

11. Knowledge retrieval (RAG). Retrieval-augmented generation: finding relevant passages in your documents and adding them to the prompt, so the agent answers from what's written down rather than from what the model happens to know. In LibraOS, knowledge retrieves from your connected sources.

12. Embedding. A list of numbers that represents the meaning of a piece of text, produced by an embedding model. Retrieval works by finding passages whose embeddings are close to the question's.

13. Grounded answer. An answer based on retrieved sources, ideally with a reference the reader can check. A citation shows where an answer came from, not that it's correct.

Control and safety

14. Permissions. What an agent is allowed to do, and with which tools. Give each agent only the tools it needs.

15. Approval point. A place where a task stops for a person's decision, such as before a refund, a message to a customer or a record change. In LibraOS, you configure permissions, approval points and audit for the tasks your app runs through the Runtime.

16. Human-in-the-loop. A design where the agent drafts and a person decides. The agent reads and proposes freely; actions with side effects wait for approval.

17. Audit trail. A structured record of a task from request to result: who asked, what context was used, which tools were called, who approved, and what changed. It lets someone who wasn't there explain what happened.

18. Prompt injection. Instructions hidden in input, such as a document or web page, that try to make the agent ignore its rules or take an action it shouldn't.

19. AI firewall. Screening for requests and responses. In LibraOS it has three tiers: pattern screening for configured injection patterns and sensitive-data formats, risk scoring with a classifier, and model review by a judge model for ambiguous cases. Screening reduces risk; it does not guarantee that every harmful prompt or response will be detected.

Models and deployment

20. Chat model. The model that writes the responses your users see. Most agents also need an embedding model for knowledge retrieval.

21. Bring your own key (BYOK). Connecting your own provider's OpenAI-compatible endpoint with your own credentials. You choose the models and manage billing with your provider.

22. Local inference. Running a compatible model server on your own machine or private network, alongside the runtime, so prompts and data stay inside it.

23. Hosted inference. Using a model service someone else runs. For LibraOS, that's Gateway: MegaNova's OpenAI-compatible model APIs, with account, access, billing and limits separate from LibraOS licensing.

24. Self-hosted. Running the runtime on infrastructure you control, on-premise or in a private cloud, with models, knowledge stores and services placed where your requirements allow.

25. Air-gapped. A deployment in a network with no internet connection. It uses local models and connectors, and every network dependency, log and update path has to be validated, with downloads verified outside before they're brought in.

Conclusion

A shared vocabulary saves time in every agent project. Agree on what you mean by memory versus knowledge, approval points versus UI buttons, and self-hosted versus air-gapped before you estimate the work, and most of the confusion in planning and reviews goes away. Keep this list next to your design doc, and add your team's own terms as they come up.

Explore LibraOS: libraos.com

Sources

• LibraOS homepage: Runtime
• Security & Private Deployment
• Models & Inference
• Developers

What’s Next?

Sign up and explore now.

🔍 Learn more: Visit our blog and documents for more insights or contact us for a demo of how MegaNova can power your AI products.

📬 Get in touch: Join our Discord community for help or Contact Us.