Glossary

• Model. A large language model. It takes text in and returns text out.
• Harness. The code around a model. It sends requests, runs tools, keeps state, and decides when to stop.
• Agent. A model plus a harness plus a goal.
• Tool. A Java method the model can ask the harness to run.
• Tool call. The model’s request to run a tool. It names the tool and gives arguments.
• Loop. The cycle of model call, tool execution, and result. It ends when the run stops.
• Termination. The rule that ends a run.
• Context window. The text the model can read in one call. It is measured in tokens.
• Token. A small piece of text. Providers bill by token.
• Skill. A file of instructions for one task. The harness loads it only when needed.
• Sub-agent. An agent that a parent agent calls as a tool.
• Memory. Information kept across model calls. Short-term memory is the recent conversation. Long-term memory is stored outside it.
• Embedding. A list of numbers that represents the meaning of a text.
• Vector store. A database that finds texts with similar embeddings.
• BM25. A keyword scoring method. It rewards rare words that appear often in a document.
• Hybrid search. Keyword search and vector search, merged into one result list.
• RRF (Reciprocal Rank Fusion). A way to merge ranked lists by rank position, not by score.
• Reranker. A model that scores a query and each candidate together, then reorders the candidates.
• RAG (retrieval-augmented generation). The harness fetches relevant documents before the model answers.
• HITL (human-in-the-loop). A run that pauses until a person approves an action.
• Graph. A flow of nodes and edges. A knowledge graph stores relationships between things.
• Durable execution. A run that survives crashes and restarts because its state is stored.
• Workflow (Temporal). The durable definition of one run.
• Activity (Temporal). A unit of work with side effects, such as a model call. Temporal retries it.
• Signal (Temporal). A message sent to a running workflow, such as an approval.
• Eval. A fixed set of test inputs with expected properties. Run it before each change.
• Prompt caching. The provider reuses work for a stable part of the prompt. It lowers cost.