| tokenomics / token economics | 토큰 이코노믹스 / 토큰 경제학 | The discipline of measuring, pricing, budgeting, and optimizing token usage, caching, and context across LLM requests and AI agent loops without degrading output quality. |
| token | 토큰 | Unit of text a model reads or writes; the basis of usage billing. |
| input / output tokens | 입력 / 출력 토큰 | Tokens sent to the model / generated by it. |
| context window | 컨텍스트 윈도우 | Maximum tokens a model can consider in one request. |
| prefix | 접두부 | The beginning of a request; an unchanged prefix may be reused by provider caches. |
| prompt caching | 프롬프트 캐싱 | Provider reuse of an unchanged request prefix, lowering cost and latency. |
| cache hit rate | 캐시 적중률 | Share of input tokens served from the cache. |
| AI credits | AI 크레딧 | GitHub Copilot’s usage-based billing unit. |
| premium request | 프리미엄 요청 | Legacy Copilot billing unit: prompt × model multiplier. |
| compaction | 컨텍스트 압축 | Summarizing older conversation to free context space. |
| observation masking | 관찰 마스킹 | Replacing older tool results with compact placeholders while retaining recent observations. |
tokenizer encoding (o200k_base / cl100k_base) | 토크나이저 인코딩 | The tokenization vocabulary and rules used to map text to tokens; counts depend on the encoding. |
| tool definition | 도구 정의 | The schema of a tool sent to the model on every request. |
| tool pruning | 도구 가지치기 | Leaving unused tool definitions out of requests. |
| micro-skill | 마이크로 스킬 | A short, confirmed repository fact stated at session start. |
| failure loop | 실패 루프 | A turn repeatedly hitting the same failure. |
| loop guard | 루프 가드 | Mechanism that detects a failure loop and redirects the agent once. |
| turn | 턴 | The agent’s work on one user prompt. |
| baseline | 기준선 | The measured starting point to compare against. |
| subagent | 서브에이전트 | A separate agent run that keeps exploration out of the main conversation. |
| context rot | 컨텍스트 열화 | Quality loss in long contexts even before the window is full (e.g., middle content is used less reliably). |
| landmine (instruction) | 지뢰 규칙 | A rule the agent cannot infer from code and would break if unknown; the kind worth putting in instructions. |
| tool schema tax | 도구 스키마 세금 | Tokens spent re-sending tool definitions on every request. |
| agent skill | 에이전트 스킬 | On-demand packaged knowledge or tooling that replaces always-on instructions. |
| reasoning effort | 추론 강도 | Setting that controls how many thinking tokens a model spends per request. |
| fan-out (LLM calls per request) | 팬아웃 (요청당 LLM 호출 수) | Number of model calls triggered by one user request in a pipeline or agent. |
| token map | 토큰 지도 | Per-step table of fixed and dynamic tokens, calls, and share of the total. |
| cacheable prefix | 캐시 가능 접두부 | The stable start of a request that meets the provider’s minimum length and can be served from the prompt cache. |
| provisioned throughput | 프로비저닝 처리량 | Reserved model capacity billed for capacity rather than per token; optimizing it means capacity and latency, not only cost. |
