Glossary

Term Meaning

Agentic loop Call Claude, run requested tools, return results, repeat until stop_reason is end_turn stop_reason Why Claude stopped: tool_use (wants a tool), end_turn (finished), max_tokens (cut off) Coordinator The hub agent that decomposes work, delegates to subagents and aggregates results Subagent An agent spawned with its own clean context and restricted tools Task tool The Agent SDK tool a coordinator uses to spawn subagents AgentDefinition SDK configuration for a subagent type: description, prompt, tools Hook Code run at a fixed point in the loop, such as before or after a tool call fork_session Branch an independent session from a shared baseline MCP Model Context Protocol, the open standard for connecting tools and data to Claude MCP tool / resource An action the model can call / content the client can read isError MCP flag marking a tool result as a failure

.mcp.json Project-scoped MCP server configuration, committed to the repo tool_choice auto, any, or a named tool to force

CLAUDE.md Always-loaded instructions for Claude Code at user, project or directory level .claude/rules/ Topic rule files, optionally scoped to file globs with paths Skill On-demand workflow defined in SKILL.md with frontmatter

Plan mode Claude Code mode that explores and plans without changing files Explore subagent Built-in subagent that isolates verbose discovery -p / --print Non-interactive Claude Code for scripts and CI

Message Batches API Asynchronous API at 50% cost, up to 24 hours, matched by custom_id Prompt chaining A fixed sequence of focused prompts

Attention dilution Uneven, inconsistent quality when one pass covers too much Lost in the middle Weaker recall of material in the middle of long inputs Case facts block Persistent structured facts kept outside summarised history Claim-source mapping A claim tied to its excerpt, source and date Stratified sampling Random review samples drawn from each segment to measure error rates