Glossary

Term Meaning

Adaptive thinking Thinking mode where Claude decides how much to reason; set with thinking: {type: "adaptive"}

Agent A system where the model directs its own tool use in a loop until the task is done Agent SDK Claude Code's harness packaged as a library, with built-in tools, hooks, and subagents; selfhosted Batch API Asynchronous endpoint for large jobs at 50% cost; results keyed by custom_id Cache breakpoint A cache_control marker that ends a cacheable prefix; up to 4 per request CLAUDE.md Memory and instruction file read by Claude Code at user, project, and folder levels Compaction Summarizing earlier conversation to free context near the limit Context drift An agent gradually losing track of earlier goals or constraints over a long session Context editing Removing specific old content, such as tool results, from the conversation Effort Setting (low to max) that trades reasoning depth for tokens and latency Eval A fixed test set and grading method run after every change Few-shot (multishot)Including several examples in the prompt

Hook Deterministic code that runs at a lifecycle event, such as before a tool call Managed Agents Anthropic-hosted agent loop and sandbox with versioned agent configs MCP Model Context Protocol: open standard for exposing tools, resources, and prompts to AI apps Prompt injection Instructions hidden in content the model reads, aiming to override its real instructions Skill A folder with SKILL.md and resources that Claude loads when relevant stop_reason Why generation ended: end_turn, max_tokens, stop_sequence, tool_use, pause_turn, refusal Streamable HTTP MCP transport for remote, multi-client servers

stdio MCP transport where the client runs the server as a local subprocess Structured outputs API feature constraining responses to a JSON schema Subagent A helper agent with its own context, tools, and prompt, returning a summary Tool Runner SDK helper that runs the tool loop for tools you define Workflow A system where code, not the model, orchestrates LLM calls through predefined steps