Glossary

Agent: A language model running in a loop. On each step, it either calls a tool or gives a final answer.

Agent loop: The code that sends messages to the model, runs the tools it asks for, sends back the results, and repeats until the model answers or a step limit is reached.

Chat completions API: The HTTP API most model servers use: you POST a list of messages and receive the next message. Originally from OpenAI, now an informal standard.

Context (context window): Everything the model can see on a given call: the system prompt, the conversation, and tool results. The model has no other memory.

Defense in depth: Using several independent layers of protection, so one failure doesn’t mean a breach.

Eval (evaluation): A test for an AI system. An input, plus one or more checks that decide whether the output is good enough.

Excessive agency: Giving an agent more access or autonomy than it needs, so a mistake or attack can do more damage.

Human in the loop: Requiring a person to approve an action before the agent carries it out.

Indirect prompt injection: Instructions hidden in data the agent reads, such as a web page, a document or a session abstract, rather than typed by the user.

Least privilege: Giving a component only the access it needs for its task, and nothing more.

LLM-as-judge: Using a language model to grade another model’s output against a rubric.

Prompt injection: Text that makes a model ignore its instructions and follow the attacker’s instead.

Rubric: A written description of what a good answer looks like, used by a judge (human or model).

System prompt: Instructions from the developer at the start of the conversation. They guide the model, but they aren’t secret and aren’t enforced.

Temperature: A setting that controls randomness. Lower values give more predictable output. Higher values give more varied output.

Tool (function) calling: A model’s ability to request that your code run a function with specific arguments. The model only asks. Your code decides whether to run it.

Tool schema: A JSON Schema description of a tool’s name, purpose and parameters, sent to the model so it knows how to call the tool.