Glossary
The words the writing here leans on without stopping to explain them. Each definition is taken from the article on this site that does explain it, and links back to it, so nothing in this list is a claim the writing does not already make.
- AI agent
An AI agent is an LLM-driven system that operates in a loop: plan, decompose a goal into sub-tasks; act, select and invoke one or more tools; observe, read the tool results and update internal state; iterate, repeat until the goal is met or a stopping condition is reached. What makes it an agent rather than a workflow is that the model decides the control flow at runtime: which tool to call, whether to loop again, when to stop.
Explained in AI Agents, MCP, and the Modern Framework Landscape, Is It Actually an Agent? I Built the Same Task Three Ways to Find Out
- Context window
How much text a model can be handed in one call. It is a ceiling rather than a slope: past it the prompt does not degrade, it fails, with zero model calls and an error saying the prompt is too long. An architecture that puts the whole dataset in the prompt therefore does not get worse as the data grows, it disappears.
Explained in Is It Actually an Agent? I Built the Same Task Three Ways to Find Out
- Control plane
The one place where a fleet of agents is inventoried, observed, governed and stoppable. "Halt every agent that touched the billing tool this hour, and tell me who approved that" is a registry question, an observability question, a policy question and an audit question at once, and a control plane is the place where those four can be one query.
Explained in I Built an AI Agent Control Plane. Here Is What Governance Actually Takes
- Eval suite
Automated, re-runnable test suites that measure an agent's success rate per tool: ten to twenty realistic tasks with expected ground-truth outputs, re-run after every fix, capturing success rate, tool-call count, token usage, latency and error rate by category.
Explained in AI Agents, MCP, and the Modern Framework Landscape
- Hallucination
The industry's own word for output that is confidently wrong: material inaccuracies that appear accurate because of their level of detail or specificity. It is the defining risk of the product, and every vendor's terms hand the duty of verifying the output to you.
Explained in Who Pays When the AI Is Wrong? I Read Ten Vendors' Terms to Find Out
- Kill switch
The control that halts agents, globally or by environment, team or agent. It has to stop new work, not just old: halting every agent is trivial to implement as a status update on existing rows and completely useless if a halted agent can open a new run a second later, so the switch has to gate the creation of work.
Explained in I Built an AI Agent Control Plane. Here Is What Governance Actually Takes
- LLMLarge Language Model
LLMs harness deep learning algorithms and neural networks, specifically transformer models, to understand and generate natural language text. Trained on vast datasets such as books, articles and websites, these models are capable of producing human-like text.
Explained in Imagine a World Where Books Answer Your Questions
- MCPModel Context Protocol
An open standard, originally created by Anthropic and now governed by the Linux Foundation. Clients, among them Claude, ChatGPT, VS Code, the JetBrains IDEs and Cursor, attach to MCP servers that expose tool schemas. Its key value: write your tools once and expose them to any MCP-compatible client, with no per-client integration work.
Explained in AI Agents, MCP, and the Modern Framework Landscape
- Observability
The record of what an agent actually did at runtime: runs, spans, model calls, tool calls, tokens, latency and cost. Without it, everything between the request and the answer is invisible, which is the first gap teams shipping agents converge on.
Explained in I Built an AI Agent Control Plane. Here Is What Governance Actually Takes
- RAGRetrieval-Augmented Generation
Connecting a language model to external data sources, such as PDFs, APIs and databases, so that the answer comes back out of those sources instead of out of the model's general training alone. The crucial part is pinpointing the tasks or questions that demand domain expertise or information beyond the model's general knowledge, and understanding what specific information is required to generate precise responses.
Explained in Imagine a World Where Books Answer Your Questions
- Tool (for an agent)
Tools are semantic contracts. An agent reads a tool's name, description and schema to decide whether and how to call it, so if any of those signals is vague the agent may call the wrong tool, pass malformed inputs, or misinterpret outputs and hallucinate downstream.
Explained in AI Agents, MCP, and the Modern Framework Landscape
- Workflow
Workflows orchestrate LLMs through predefined code paths; agents direct their own process. A fixed classify, execute, compose chain is a workflow however many LLM calls it contains, because every path it can take was enumerated by a human at author time.
Explained in Is It Actually an Agent? I Built the Same Task Three Ways to Find Out
Last updated