[{"data":1,"prerenderedAt":838},["ShallowReactive",2],{"blog-post-ai-agent-glossary":3,"related-posts-ai-agent-glossary":837},{"id":4,"title":5,"author":6,"body":10,"category":816,"date":817,"description":818,"extension":819,"featured":820,"image":821,"imageHeight":822,"imageWidth":822,"meta":823,"navigation":824,"path":825,"readingTime":826,"seo":827,"seoTitle":828,"stem":829,"tags":830,"updatedDate":817,"__hash__":836},"blog/blog/ai-agent-glossary.md","AI Agent Glossary: 50 Terms Explained Without Buzzwords",{"name":7,"role":8,"avatar":9},"Shabnam Katoch","Growth Head","/img/avatars/shabnam-profile.jpeg",{"type":11,"value":12,"toc":803},"minimark",[13,17,20,23,28,31,72,109,146,187,208,241,274,277,281,285,293,300,303,306,309,316,319,322,325,328,331,334,337,344,347,350,352,356,359,365,371,374,377,380,383,386,389,392,395,398,401,404,407,410,421,423,427,430,438,444,447,450,453,459,462,465,468,471,474,483,486,489,492,498,500,504,507,515,521,524,530,533,539,542,548,551,559,562,565,568,571,574,582,585,588,590,594,597,605,611,614,620,623,629,632,642,644,648,651,657,660,667,670,673,676,679,682,685,688,691,694,700,702,706,709,715,721,724,730,733,743,746,749,752,755,758,773,776,779,781,785,788,794,797],[14,15,16],"p",{},"If you have spent any time around AI agents in the last year, you have hit the wall. Someone drops a term like \"context window\" or \"MCP\" or \"agentic loop\" into a sentence and just keeps going, assuming you already know what it means. Most glossaries do not help. They are either written for people with a computer science degree or stuffed with so much marketing language that you finish a definition knowing less than when you started.",[14,18,19],{},"This is the clean one. Fifty terms, each explained in two or three plain sentences, no jargon hiding behind more jargon. Every term has its own anchor link so you can bookmark the exact definition you need or point a teammate straight to it. We verified the moving pieces (protocol versions, current standards, recent launches) against primary sources as of June 2026, because half the AI glossaries online are quietly out of date.",[14,21,22],{},"Bookmark this page. We update it as the space moves.",[24,25,27],"h2",{"id":26},"how-to-use-this-glossary","How to use this glossary",[14,29,30],{},"Terms are grouped into seven sections so related concepts sit next to each other: the basics, how agents think and act, memory and context, tools and protocols, multi agent systems, building and running agents, and safety and trust. Jump to any term using the links below.",[14,32,33,37,38,43,44,43,48,43,52,43,56,43,60,43,64,43,68],{},[34,35,36],"strong",{},"The basics:"," ",[39,40,42],"a",{"href":41},"#ai-agent","AI Agent"," · ",[39,45,47],{"href":46},"#llm","LLM",[39,49,51],{"href":50},"#agentic","Agentic",[39,53,55],{"href":54},"#autonomy","Autonomy",[39,57,59],{"href":58},"#inference","Inference",[39,61,63],{"href":62},"#token","Token",[39,65,67],{"href":66},"#prompt","Prompt",[39,69,71],{"href":70},"#system-prompt","System Prompt",[14,73,74,37,77,43,81,43,85,43,89,43,93,43,97,43,101,43,105],{},[34,75,76],{},"How agents think and act:",[39,78,80],{"href":79},"#agentic-loop","Agentic Loop",[39,82,84],{"href":83},"#reasoning","Reasoning",[39,86,88],{"href":87},"#planning","Planning",[39,90,92],{"href":91},"#react","ReAct",[39,94,96],{"href":95},"#chain-of-thought","Chain of Thought",[39,98,100],{"href":99},"#tool-calling","Tool Calling",[39,102,104],{"href":103},"#function-calling","Function Calling",[39,106,108],{"href":107},"#orchestration","Orchestration",[14,110,111,37,114,43,118,43,122,43,126,43,130,43,134,43,138,43,142],{},[34,112,113],{},"Memory and context:",[39,115,117],{"href":116},"#context-window","Context Window",[39,119,121],{"href":120},"#context-engineering","Context Engineering",[39,123,125],{"href":124},"#memory","Memory",[39,127,129],{"href":128},"#rag","RAG",[39,131,133],{"href":132},"#agentic-rag","Agentic RAG",[39,135,137],{"href":136},"#vector-database","Vector Database",[39,139,141],{"href":140},"#embedding","Embedding",[39,143,145],{"href":144},"#chunking","Chunking",[14,147,148,37,151,43,155,43,159,43,163,43,167,43,171,43,175,43,179,43,183],{},[34,149,150],{},"Tools and protocols:",[39,152,154],{"href":153},"#mcp","MCP",[39,156,158],{"href":157},"#mcp-server","MCP Server",[39,160,162],{"href":161},"#a2a","A2A",[39,164,166],{"href":165},"#agent-card","Agent Card",[39,168,170],{"href":169},"#skill","Skill",[39,172,174],{"href":173},"#integration","Integration",[39,176,178],{"href":177},"#oauth","OAuth",[39,180,182],{"href":181},"#byok","BYOK",[39,184,186],{"href":185},"#api","API",[14,188,189,37,192,43,196,43,200,43,204],{},[34,190,191],{},"Multi agent systems:",[39,193,195],{"href":194},"#multi-agent-system","Multi Agent System",[39,197,199],{"href":198},"#sub-agent","Sub Agent",[39,201,203],{"href":202},"#handoff","Handoff",[39,205,207],{"href":206},"#agent-mesh","Agent Mesh",[14,209,210,37,213,43,217,43,221,43,225,43,229,43,233,43,237],{},[34,211,212],{},"Building and running agents:",[39,214,216],{"href":215},"#no-code-agent-builder","No Code Agent Builder",[39,218,220],{"href":219},"#agent-framework","Agent Framework",[39,222,224],{"href":223},"#deployment","Deployment",[39,226,228],{"href":227},"#cron","Cron",[39,230,232],{"href":231},"#webhook","Webhook",[39,234,236],{"href":235},"#trigger","Trigger",[39,238,240],{"href":239},"#fine-tuning","Fine Tuning",[14,242,243,37,246,43,250,43,254,43,258,43,262,43,266,43,270],{},[34,244,245],{},"Safety and trust:",[39,247,249],{"href":248},"#hallucination","Hallucination",[39,251,253],{"href":252},"#guardrails","Guardrails",[39,255,257],{"href":256},"#trust-level","Trust Level",[39,259,261],{"href":260},"#kill-switch","Kill Switch",[39,263,265],{"href":264},"#sandboxing","Sandboxing",[39,267,269],{"href":268},"#prompt-injection","Prompt Injection",[39,271,273],{"href":272},"#human-in-the-loop","Human in the Loop",[275,276],"hr",{},[24,278,280],{"id":279},"the-basics","The basics",[282,283,42],"h4",{"id":284},"ai-agent",[14,286,287,288,292],{},"An ",[39,289,291],{"href":290},"/blog/what-is-ai-agent","AI agent"," is a program that uses a large language model to decide what to do, then actually does it by calling tools, reading data, and taking actions on your behalf. The difference between an agent and a regular chatbot is action: a chatbot answers, an agent gets things done. Think of it as the difference between an assistant who gives you advice and one who actually books the meeting.",[14,294,295],{},[296,297],"img",{"alt":298,"src":299},"Split illustration contrasting a chatbot that only answers with an AI agent that completes a checklist","/img/blog/ai-agent-glossary-chatbot-vs-agent.jpg",[282,301,47],{"id":302},"llm",[14,304,305],{},"LLM stands for large language model. It is the underlying brain that powers an AI agent, trained on huge amounts of text to predict and generate language, which lets it answer questions, write, and reason. Examples you have heard of include the models from OpenAI, Anthropic, and Google, and an agent typically sits on top of one of them.",[282,307,51],{"id":308},"agentic",[14,310,311,312,315],{},"\"",[39,313,51],{"href":314},"/blog/what-is-agentic-ai","\" just means an AI system that can act on its own toward a goal rather than waiting for instructions at every step. If a tool can decide its next move, use other tools, and keep going until a task is done, people call it agentic. It is the adjective everyone reaches for when \"chatbot\" no longer fits.",[282,317,55],{"id":318},"autonomy",[14,320,321],{},"Autonomy is how much an agent is allowed to do without checking back with a human first. A low autonomy agent drafts an email and waits for you to hit send, while a high autonomy agent sends it and moves on to the next task. Most real deployments pick a level in between, depending on how much a mistake would cost.",[282,323,59],{"id":324},"inference",[14,326,327],{},"Inference is the moment a model actually runs and produces an output from your input. Every time an agent \"thinks,\" it is running inference, and each run costs money and time. When people talk about inference costs, they mean the bill you pay your model provider for all that thinking.",[282,329,63],{"id":330},"token",[14,332,333],{},"A token is a chunk of text, usually a word or part of a word, and it is the unit models read and write in. Both your input and the model's output are measured in tokens, which is also how you get billed. A rough rule: one token is about four characters of English, so this sentence is roughly fifteen tokens.",[282,335,67],{"id":336},"prompt",[14,338,339,340,343],{},"A prompt is the text instruction you give a model to tell it what you want. It can be a single question or a detailed set of directions, and the quality of the prompt heavily shapes the quality of the answer. For agents, prompts are less about clever wording and more about giving the right information, which is its own discipline covered under ",[39,341,342],{"href":120},"context engineering",".",[282,345,71],{"id":346},"system-prompt",[14,348,349],{},"A system prompt is the hidden set of instructions that defines who an agent is and how it should behave, set once and applied to every conversation. It is where you tell the agent its role, its limits, and its tone before any user ever talks to it. Think of it as the job description an agent reads before clocking in.",[275,351],{},[24,353,355],{"id":354},"how-agents-think-and-act","How agents think and act",[282,357,80],{"id":358},"agentic-loop",[14,360,361,362,364],{},"The agentic loop is the cycle an agent repeats to get work done: look at the situation, decide on an action, take it, check the result, and repeat until the goal is met. This loop is what separates an agent from a single question and answer, because the agent keeps going on its own. Every step generates new information the agent has to fit into its limited working memory, which is why ",[39,363,342],{"href":120}," matters so much.",[14,366,367],{},[296,368],{"alt":369,"src":370},"Circular agentic loop diagram showing the observe, decide, act, and check stages","/img/blog/ai-agent-glossary-agentic-loop.jpg",[282,372,84],{"id":373},"reasoning",[14,375,376],{},"Reasoning is the model's ability to work through a problem in steps instead of jumping straight to an answer. Newer models can spend extra effort \"thinking\" before they respond, which improves accuracy on hard, multi step tasks. For agents, better reasoning usually means better decisions about which tool to use and when.",[282,378,88],{"id":379},"planning",[14,381,382],{},"Planning is when an agent breaks a big goal into a sequence of smaller steps before acting on any of them. Instead of reacting one move at a time, the agent maps out a route, which helps on complex tasks with many moving parts. A travel agent that books flights, hotels, and a rental car in order is planning, not just reacting.",[282,384,92],{"id":385},"react",[14,387,388],{},"ReAct is short for \"reason and act,\" a common pattern where an agent alternates between thinking out a step and then taking an action based on that thinking. The agent reasons about what to do, does it, observes what happened, and reasons again. It is one of the foundational patterns behind how most modern agents operate.",[282,390,96],{"id":391},"chain-of-thought",[14,393,394],{},"Chain of thought is a technique where a model is encouraged to show its step by step reasoning rather than leaping to a conclusion. Writing out the intermediate steps tends to produce more accurate results on math, logic, and multi part problems. You can think of it as making the model \"show its work\" the way a teacher asks a student to.",[282,396,100],{"id":397},"tool-calling",[14,399,400],{},"Tool calling is when an agent uses an external capability, like searching the web, querying a database, or sending a message, instead of relying only on what the model already knows. The model decides a tool is needed, picks the right one, and passes in the inputs. This is the mechanism that turns a language model from a talker into a doer.",[282,402,104],{"id":403},"function-calling",[14,405,406],{},"Function calling is the technical name for the same idea as tool calling, often used when the tool is a specific piece of code the model can invoke. The model returns a structured request saying \"call this function with these inputs,\" and your system runs it. In practice, people use \"tool calling\" and \"function calling\" almost interchangeably.",[282,408,108],{"id":409},"orchestration",[14,411,412,415,416,420],{},[39,413,108],{"href":414},"/blog/ai-agent-orchestration"," is the coordination layer that decides which tools, models, or other agents handle each part of a task and in what order. When a workflow involves several steps and several capabilities, something has to route the work, and that something is the orchestrator. On a no code platform like ",[39,417,419],{"href":418},"/","BetterClaw",", this routing happens for you behind a visual builder instead of in code.",[275,422],{},[24,424,426],{"id":425},"memory-and-context","Memory and context",[282,428,117],{"id":429},"context-window",[14,431,432,433,437],{},"The ",[39,434,436],{"href":435},"/blog/ai-agent-context-window-explained","context window"," is the maximum amount of text, measured in tokens, that a model can consider at one time, covering both your input and its output. Once a conversation outgrows the window, older parts get dropped or summarized, which is why long running agents need careful memory handling. Bigger windows help, but cramming more in is not the same as using it well.",[14,439,440],{},[296,441],{"alt":442,"src":443},"Sliding context window over a long scroll of text with the visible portion highlighted","/img/blog/ai-agent-glossary-context-window.jpg",[282,445,121],{"id":446},"context-engineering",[14,448,449],{},"Context engineering is the practice of deciding exactly what information makes it into the model's limited working memory at each step, and in what form. It is broader than writing a good prompt, covering memory, retrieved documents, tool outputs, and how all of it gets compressed and ordered. The industry has shifted hard toward this in 2026, with most data leaders agreeing that clever prompts alone no longer cut it in production.",[282,451,125],{"id":452},"memory",[14,454,455,458],{},[39,456,125],{"href":457},"/blog/how-ai-agent-memory-works"," is how an agent holds onto information across turns or across sessions so it does not start from scratch every time. Short term memory lives inside the current conversation, while long term memory persists facts, preferences, and past events for later. Without memory, an agent forgets your name the moment the context window fills up.",[282,460,129],{"id":461},"rag",[14,463,464],{},"RAG stands for retrieval augmented generation. It is the technique of pulling relevant documents or data into the model's context right before it answers, so the response is grounded in real, current information instead of only what the model memorized during training. It is how an agent can answer questions about your private docs or last week's data.",[282,466,133],{"id":467},"agentic-rag",[14,469,470],{},"Agentic RAG is RAG where the agent itself decides when to retrieve, what to search for, and whether the results are good enough, instead of always grabbing documents up front. A plain RAG system retrieves once and answers, while an agentic version can search, evaluate, and search again. This makes it better at messy questions where the first search rarely finds everything.",[282,472,137],{"id":473},"vector-database",[14,475,476,477,479,480,482],{},"A vector database is a specialized store that holds text as numerical representations called embeddings, so an agent can find information by meaning rather than exact keywords. When an agent needs to \"remember\" or retrieve something relevant, it searches this database for the closest matches. It is the storage layer that makes ",[39,478,129],{"href":128}," and long term ",[39,481,452],{"href":124}," practical at scale.",[282,484,141],{"id":485},"embedding",[14,487,488],{},"An embedding is a list of numbers that captures the meaning of a piece of text, so that similar ideas end up with similar numbers. This is what lets a system find \"refund policy\" when you searched for \"how do I get my money back,\" even though the words differ. Embeddings are the math trick underneath semantic search and most agent memory.",[282,490,145],{"id":491},"chunking",[14,493,494,495,497],{},"Chunking is the practice of splitting a long document into smaller pieces before storing it, so retrieval can pull just the relevant section instead of the whole file. Chunk too big and you waste context on irrelevant text, chunk too small and you lose the surrounding meaning. Getting chunk size right is one of the quiet but important decisions in any ",[39,496,129],{"href":128}," setup.",[275,499],{},[24,501,503],{"id":502},"tools-and-protocols","Tools and protocols",[282,505,154],{"id":506},"mcp",[14,508,509,510,514],{},"MCP stands for ",[39,511,513],{"href":512},"/blog/what-is-mcp-model-context-protocol","Model Context Protocol",". It is an open standard for how an agent connects to external tools and data sources, so any compliant tool plugs into any compliant agent without custom code for each pairing. Think of it as USB for AI: you plug a tool in and it just works. It launched in late 2024, became the de facto standard within a year, and is now governed under the Linux Foundation after Anthropic donated it in December 2025.",[14,516,517],{},[296,518],{"alt":519,"src":520},"An AI agent shown as a central hub with external tools plugging in like USB devices through MCP","/img/blog/ai-agent-glossary-mcp-usb.jpg",[282,522,158],{"id":523},"mcp-server",[14,525,526,527,529],{},"An MCP server is the piece that wraps a specific tool or data source and exposes it over ",[39,528,154],{"href":153}," so agents can use it. If MCP is the universal plug, an MCP server is the device on the other end, like one for your email, your calendar, or your database. There are now thousands of them, and platforms increasingly let you connect one with a single click.",[282,531,162],{"id":532},"a2a",[14,534,535,536,538],{},"A2A stands for Agent2Agent, an open protocol for how separate AI agents talk to each other, even when they were built by different teams on different frameworks. Where ",[39,537,154],{"href":153}," connects an agent to tools, A2A connects an agent to other agents so they can delegate and coordinate. Introduced by Google in 2025 and now under the Linux Foundation, it reached a production grade v1.0 in early 2026 with signed agent cards for verifying who is who.",[282,540,166],{"id":541},"agent-card",[14,543,544,545,547],{},"An agent card is a small metadata document that describes what an agent can do and how to reach it, used by ",[39,546,162],{"href":161}," so agents can discover each other. It is like a business card for an agent, listing its skills and its address in a standard format. As of the v1.0 standard, these cards can be cryptographically signed so a receiving agent can confirm the card is genuine and not a forgery.",[282,549,170],{"id":550},"skill",[14,552,553,554,558],{},"A skill is a packaged capability you can add to an agent, like the ability to read Gmail, post to Slack, or check Google Search Console, usually defined in a standard file format. Skills are how you extend what an agent can do without rebuilding it from scratch. BetterClaw ships ",[39,555,557],{"href":556},"/blog/best-openclaw-skills","200+ verified skills"," that pass a security audit before they reach your agent.",[282,560,174],{"id":561},"integration",[14,563,564],{},"An integration is a connection between your agent and an outside app or service, such as your CRM, your calendar, or your messaging tool. Each integration lets the agent read from or act inside that service. The practical difference between platforms often comes down to how many integrations they offer and how hard each one is to set up.",[282,566,178],{"id":567},"oauth",[14,569,570],{},"OAuth is the secure standard that lets you grant an agent access to an app, like your Google account, without ever handing over your password. You log in on the real service, approve specific permissions, and the agent receives a limited token instead of your credentials. It is the \"sign in with Google\" flow you have clicked a hundred times, working quietly behind agent integrations.",[282,572,182],{"id":573},"byok",[14,575,576,577,581],{},"BYOK stands for bring your own key. It means you supply your own API key for the AI model provider you want to use, so you pay the provider directly at their normal rates instead of paying a marked up price through the platform. It gives you cost transparency and the freedom to switch providers anytime. BetterClaw's ",[39,578,580],{"href":579},"/pricing","free plan"," is BYOK with zero markup on inference.",[282,583,186],{"id":584},"api",[14,586,587],{},"API stands for application programming interface, a defined way for one piece of software to talk to another. Agents use APIs constantly, both to reach the underlying model and to act inside the apps they integrate with. When someone says an agent \"calls the API,\" they mean it is sending a structured request to another service and getting a structured answer back.",[275,589],{},[24,591,593],{"id":592},"multi-agent-systems","Multi agent systems",[282,595,195],{"id":596},"multi-agent-system",[14,598,599,600,604],{},"A ",[39,601,603],{"href":602},"/blog/openclaw-multi-agent-setup","multi agent system"," is a setup where several specialized agents work together on a problem that would be too much for one agent alone. Instead of one generalist trying to do everything, you get a team where each agent handles what it is good at. A research agent, a writing agent, and a fact checking agent passing work between them is a simple example.",[14,606,607],{},[296,608],{"alt":609,"src":610},"Three specialized AI agents passing a task between them in a multi agent system","/img/blog/ai-agent-glossary-multi-agent.jpg",[282,612,199],{"id":613},"sub-agent",[14,615,616,617,619],{},"A sub agent is a smaller agent that a main agent spins up to handle a specific piece of a larger task. The main agent stays focused on the overall goal and delegates the narrow work, then folds the result back in. This keeps each agent's ",[39,618,436],{"href":116}," clean instead of one agent juggling everything at once.",[282,621,203],{"id":622},"handoff",[14,624,625,626,628],{},"A handoff is the moment one agent passes a task, along with the relevant context, to another agent better suited to finish it. A support agent that transfers a billing question to a billing agent mid conversation is performing a handoff. Doing this cleanly, without losing the thread, is one of the hard problems that protocols like ",[39,627,162],{"href":161}," were built to solve.",[282,630,207],{"id":631},"agent-mesh",[14,633,634,635,637,638,641],{},"An agent mesh is a network of agents that can discover and talk to each other directly, rather than all routing through one central controller. The term describes the architecture you get when many agents, possibly from different vendors, coordinate as peers. It is the larger vision that ",[39,636,162],{"href":161}," and signed ",[39,639,640],{"href":165},"agent cards"," are meant to enable.",[275,643],{},[24,645,647],{"id":646},"building-and-running-agents","Building and running agents",[282,649,216],{"id":650},"no-code-agent-builder",[14,652,653,654,343],{},"A no code agent builder is a platform that lets you create and deploy agents through a visual interface, without writing code, Docker configs, or YAML files. You describe what you want in plain English, connect your integrations, and deploy. This is the category BetterClaw sits in, with a visual builder that gets an agent running in about ",[39,655,656],{"href":418},"60 seconds",[282,658,220],{"id":659},"agent-framework",[14,661,662,663,666],{},"An agent framework is a code library that developers use to build agents from scratch, giving them the building blocks but leaving the assembly to them. Frameworks offer maximum flexibility at the cost of requiring real engineering time. They are the opposite end of the spectrum from a ",[39,664,665],{"href":215},"no code builder",", and which one fits depends on whether you want control or speed.",[282,668,224],{"id":669},"deployment",[14,671,672],{},"Deployment is the act of putting an agent into a live environment where it actually runs and does its job, as opposed to just testing it. A deployed agent is connected, scheduled, and ready to act on real data. On managed platforms, deployment is a button, while on raw infrastructure it can mean hours of configuring servers and ports.",[282,674,228],{"id":675},"cron",[14,677,678],{},"Cron is a way to schedule an agent to run automatically at set times, like every morning at 8am or once an hour. The name comes from a long standing Unix scheduling tool, and the idea is the same: recurring tasks without you pressing go each time. A daily briefing agent that summarizes your inbox at dawn runs on a cron schedule.",[282,680,232],{"id":681},"webhook",[14,683,684],{},"A webhook is a way for an outside app to instantly notify your agent when something happens, by sending a message to a specific address. Instead of your agent constantly asking \"anything new yet?,\" the app pushes the update the moment it occurs. A new order in your store can fire a webhook that wakes up an agent to handle it.",[282,686,236],{"id":687},"trigger",[14,689,690],{},"A trigger is whatever causes an agent to start running, whether that is a schedule, an incoming message, a webhook, or a manual click. Triggers are the \"when\" of an agent, defining the events it reacts to. Setting up the right triggers is how you move from running an agent by hand to having it work on its own.",[282,692,240],{"id":693},"fine-tuning",[14,695,696,697,699],{},"Fine tuning is the process of further training a model on your own examples so it gets better at a specific task or adopts a specific style. It changes the model's internal weights, unlike ",[39,698,129],{"href":128}," or prompting, which leave the model alone and just feed it better information. It is powerful but costly, and for most agent use cases, good context beats fine tuning.",[275,701],{},[24,703,705],{"id":704},"safety-and-trust","Safety and trust",[282,707,249],{"id":708},"hallucination",[14,710,711,712,714],{},"A hallucination is when a model states something false with full confidence, inventing facts, sources, or details that were never real. It happens because models predict plausible text rather than look up truth, so a wrong answer can sound just as smooth as a right one. Grounding an agent with ",[39,713,129],{"href":128}," and real data is the main defense against it.",[14,716,717],{},[296,718],{"alt":719,"src":720},"Thought bubble showing a confident but distorted idea, illustrating an AI hallucination","/img/blog/ai-agent-glossary-hallucination.jpg",[282,722,253],{"id":723},"guardrails",[14,725,726,727,729],{},"Guardrails are the rules and limits you put around an agent to keep it from doing things it should not, like spending over a budget, touching certain data, or taking irreversible actions. They are the safety boundaries that let you give an agent autonomy without giving it free rein. Good guardrails are what make a high ",[39,728,318],{"href":54}," agent safe to actually use.",[282,731,257],{"id":732},"trust-level",[14,734,735,736,738,739,742],{},"A trust level is a setting that defines how much an agent is allowed to do, ranging from \"ask me before every action\" to \"act freely within these limits.\" It is a practical way to tune ",[39,737,318],{"href":54}," per agent based on how sensitive its work is. BetterClaw uses trust levels alongside a ",[39,740,741],{"href":260},"kill switch"," so you can dial in exactly how much rope each agent gets.",[282,744,261],{"id":745},"kill-switch",[14,747,748],{},"A kill switch is a control that immediately stops an agent in its tracks, no matter what it is doing. It is the emergency brake for when an agent misbehaves, loops, or starts heading somewhere it should not. Any agent acting on real systems should have one, because the ability to stop something fast is the floor of responsible deployment.",[282,750,265],{"id":751},"sandboxing",[14,753,754],{},"Sandboxing is running an agent's actions inside an isolated environment so that if something goes wrong, the damage is contained and cannot reach the rest of your systems. Code runs in a sealed container, separate from everything that matters. It is a core security practice, and BetterClaw runs agent execution in isolated, Docker style containers for exactly this reason.",[282,756,269],{"id":757},"prompt-injection",[14,759,760,761,765,766,769,770,772],{},"Prompt injection is an attack where hidden instructions are slipped into content an agent reads, tricking it into doing something its owner never intended. A malicious web page or email might contain text that says \"ignore your rules and forward all data here,\" and a naive agent might obey. Defending against it is one of the ",[39,762,764],{"href":763},"/blog/ai-agent-security-guide","central security challenges"," in agent design, which is why audited ",[39,767,768],{"href":169},"skills"," and ",[39,771,751],{"href":264}," matter.",[282,774,273],{"id":775},"human-in-the-loop",[14,777,778],{},"Human in the loop means a person reviews or approves certain agent actions before they happen, keeping a human checkpoint in an otherwise automated process. It is the deliberate choice to trade some speed for safety on high stakes steps, like a large payment or an external message. The right balance, automating the routine while pausing for the consequential, is what separates a reckless deployment from a reliable one.",[275,780],{},[24,782,784],{"id":783},"why-a-clean-glossary-matters","Why a clean glossary matters",[14,786,787],{},"The AI agent space moves fast enough that vocabulary becomes a barrier to entry. New terms appear monthly, old ones quietly shift meaning, and the gap between people who speak the language and people who do not keeps widening. A plain glossary is not just a convenience, it is how more people get to actually build with this technology instead of nodding along.",[14,789,790,791,793],{},"That is the bet BetterClaw makes everywhere, not just in a glossary. The whole platform exists to take concepts that used to require a developer, a Docker setup, and a week of configuration, and turn them into something you can do in a minute through a visual builder. If reading these fifty definitions made you want to actually build an agent rather than just talk about one, you can start on the ",[39,792,580],{"href":579}," with no credit card.",[14,795,796],{},"We keep this page current as the space evolves. If a term is missing or a definition has gone stale, that is on us to fix, and we will.",[14,798,799],{},[296,800],{"alt":801,"src":802},"A person building an AI agent with a friendly assistant character in a visual no code builder","/img/blog/ai-agent-glossary-closing.jpg",{"title":804,"searchDepth":805,"depth":805,"links":806},"",2,[807,808,809,810,811,812,813,814,815],{"id":26,"depth":805,"text":27},{"id":279,"depth":805,"text":280},{"id":354,"depth":805,"text":355},{"id":425,"depth":805,"text":426},{"id":502,"depth":805,"text":503},{"id":592,"depth":805,"text":593},{"id":646,"depth":805,"text":647},{"id":704,"depth":805,"text":705},{"id":783,"depth":805,"text":784},"Fundamentals","2026-06-19","Fifty AI agent terms explained in plain English, from MCP and RAG to tokens, A2A, and guardrails. Each with its own anchor link, verified current to June 2026.","md",false,"/img/blog/ai-agent-glossary.jpg",null,{},true,"/blog/ai-agent-glossary","13 min read",{"title":5,"description":818},"AI Agent Glossary: 50 Terms Explained (No Buzzwords)","blog/ai-agent-glossary",[831,832,833,834,835],"ai agent glossary","ai agent terms","ai agent terminology","what is mcp","what is rag","RYvW26GfN-VkqWEcy-ZniKt0j48MXx83yBz7CKrzvUA",[],1781872970271]