Agent
An AI system that can take actions in the world. Not just generate text but call tools, make decisions, and follow up.
Agent Loop
The core runtime cycle. The model thinks, calls a tool, reads the result, thinks again, repeats until done.
Context Window
How much information the model can hold in its head in a single call. Finite. Filling it with junk degrades quality.
Context Rot
The slow degradation of agent performance as a session gets longer and accumulates irrelevant information.
Evals
Tests that measure whether an agent actually works. Without them you cannot tell if changes made things better or worse.
Hallucination
When a model confidently makes up information. Mitigated by giving it access to real data through tools.
Harness
The software wrapper that turns a model into an agent. Includes the loop, tools, memory, and context management.
MCP
Model Context Protocol. The open standard for connecting AI agents to external systems. Anthropic created it; the industry adopted it.
Memory
Persistent storage that lets an agent remember things between sessions. Lives outside the context window.
Model
The large language model itself. Claude, GPT, Gemini. The brain of the system.
RAG
Retrieval-Augmented Generation. The pattern of pulling relevant documents into the prompt before the model answers.
Skill
A reusable instruction pack that bundles a workflow plus its supporting files. Loaded when relevant.
Subagent
A specialized agent spawned by a parent agent to handle a specific subtask in isolation.
Token
The unit a model reads and writes. Roughly three-quarters of a word in English. Used for billing.
Tool
A function the model can call to take action. The agent's hands.
Tool Call
A structured request from the model asking the harness to run a specific function with specific inputs.