AI agent glossary
This AI agent glossary defines the 16 terms used across the setups and guides, one sentence each.
- Agent loop
- A model calls tools, reads the results and repeats until it answers without a tool call.
- Unrestricted AI agent
- An agent whose model has its refusal behavior removed, run on systems the user owns or may test.
- Abliteration
- Removing the refusal direction from a model's weights so it stops refusing, without retraining. Source
- Refusal direction
- A single direction in a model's activations that mediates refusal, per Arditi et al., 2024. Source
- Base URL
- The API root a framework sends requests to. Changing it moves the agent to another provider.
- Chat Completions
- OpenAI's request format, with tools and tool_calls. Most agent frameworks speak it.
- Messages API
- Anthropic's request format, with tool_use and tool_result blocks.
- Responses API
- OpenAI's newer format. Some SDKs default to it, so set them to Chat Completions when a provider lacks it.
- Tool calling
- The model returns a structured call with a name and JSON arguments instead of plain text.
- MCP
- Model Context Protocol, an open standard for plugging tools into agents. Goose uses it for extensions.
- Checkpointer
- LangGraph's store for thread state, which gives an agent short-term memory.
- Session
- The OpenAI Agents SDK's store for conversation history across runs.
- Browser agent
- An agent that drives a real browser, such as browser-use or Hermes Agent's browser toolset.
- Cached input
- Prompt tokens a provider has seen before, billed at a lower rate. Agent loops resend history, so caching matters.
- Context window
- The most tokens a model reads in one request. Long agent runs grow toward it.
- Max output
- The most tokens a model writes in one reply. Kimi K3.1 Pro allows 943,718.