--- title: The glossary description: Every term in agentic engineering, in one searchable place. Filter by kind, or press / to jump to the search box. author: Asem Abdo course: Agentic Engineering lesson: Glossary url: https://openplanr.dev/learn/agentic-engineering/glossary updated: 2026-10-17 related: - https://openplanr.dev/learn/agentic-engineering/roadmap.md - https://openplanr.dev/learn/agentic-engineering/flashcards.md --- # The glossary By [Asem Abdo](https://www.linkedin.com/in/asemabdo/) ยท Updated 2026-10-17 *Reference* Every term in agentic engineering, in one searchable place. Filter by kind, or press / to jump to the search box. ## v - **Agent:** A model that decides its own next step and calls tools in a loop until a goal or budget stops it. - **Agent loop:** Read context, propose a tool call, run it, append the result, repeat. - **Agentic engineering:** Karpathy's February 2026 term for professional agent-driven development: you orchestrate agents that write most code and provide oversight; engineering because it takes expertise. - **Budget:** A hard cap on steps, tokens, time or cost for one run. - **Compaction:** Summarizing older conversation to free context; lossy by design. - **Context engineering:** The set of strategies for curating and maintaining the optimal set of tokens during inference (Anthropic, September 2025). - **Context rot:** Declining recall and accuracy as the number of tokens in the context window grows. - **Context window:** Everything the model can see in one call, measured in tokens. - **Eval:** A repeatable test that scores model or agent output over a set of cases. - **Extended thinking:** The model spends extra tokens reasoning before answering. - **Function calling:** Another name for tool use: the model emits a structured call for your code to run. - **Grounding:** Tying model output to retrieved facts or sources. - **Guardrail:** A check enforced in code that blocks or corrects unsafe or invalid actions. - **Hallucination:** Confident output not supported by facts or sources. - **Harness:** Everything around the model that makes it an agent: prompt, tools, permissions, hooks, memory, orchestration. - **Harness engineering:** Designing the guides and sensors that wrap a coding agent to increase confidence in its output; named as a discipline in early 2026. - **Just-in-time context:** Loading rules or documents at the moment they matter instead of up front. - **Least privilege:** Giving an agent only the tools and access the task needs. - **Lethal trifecta:** Private data, untrusted input and an exfiltration path in one agent; enables data theft by injection. - **Memory:** Facts persisted across sessions, usually in files the agent reloads. - **Non-determinism:** The same input can produce different outputs. - **Progressive disclosure:** Loading a short summary always and detail only when needed. - **Prompt engineering:** Choosing the wording of instructions given to a model. - **Prompt injection:** Untrusted content that tries to act as instructions to the agent. - **Schema validation:** Checking that structured output contains the required fields and types. - **Stop condition:** The rule that ends an agent loop. - **Structured outputs:** Responses constrained to a JSON schema. - **System prompt:** Standing instructions that frame every turn. - **Token:** The unit of text a model reads and writes; drives cost and latency. - **Tool:** A function the model may ask the harness to run, defined by name, description and input schema. - **Tool schema:** The JSON schema describing a tool's input; the model reads it to form calls. - **Builder harness:** The part of a coding agent's harness shipped by its makers: system prompt, built-in tools, loop, compaction. - **User harness:** The outer harness a team builds around a coding agent for its own codebase. - **Harness drift:** Guides and sensors contradicting each other as a harness grows; fixed by auditing the harness like code. - **Loop engineering:** Designing the goal, checks, stop conditions, state and budget so an agent works toward a verified result without a human prompting each step. An emerging term, not yet standardized. - **Vibe coding:** Karpathy's February 2025 term for accepting model-written code without closely reading it. ## t - **Agentic search:** Iterative searching with tools such as grep and file reads instead of a single retrieval step. - **Brief:** The prompt a parent agent writes for a subagent; everything the subagent will know. - **Chunking:** Splitting documents into retrievable pieces for RAG. - **Citations:** Links from claims to the retrieved sources that support them. - **CLI vs MCP:** CLI when a shell and a well-known tool exist; MCP for no-shell clients, managed auth and SaaS. - **Depth limit:** A cap on how many levels agents may delegate to other agents. - **Embeddings:** Vectors that represent the meaning of text, used for similarity search. - **Errors as observations:** Returning tool failures to the model as results so it can correct itself. - **Evaluator-optimizer:** One agent produces, another critiques, repeat until good enough. - **Fan-out / fan-in:** Running independent subtasks in parallel and combining the results. - **Handoff:** Passing a task and its context from one agent to another. - **Human in the loop:** A person approves at defined checkpoints where mistakes are expensive. - **Human review gate:** A pause in a pipeline that waits for approval before an expensive or irreversible step. - **Hybrid search:** Combining keyword and vector search for retrieval. - **Model routing:** Sending each request to a specific model or specialist based on its type. - **Model tiering:** Using strong models for planning and review, cheaper ones for execution. - **Orchestrator-worker:** A lead agent splits work, workers execute, the lead merges. - **Prompt chaining:** A fixed sequence of model calls where each output feeds the next. - **RAG:** Retrieval-augmented generation: retrieve relevant documents, add them to the prompt, then generate. - **Reranking:** Reordering retrieved results by relevance with a second, more precise model. - **Verifier agent:** An agent that re-checks claims against independent evidence. - **Workflow:** Fixed steps defined in code, with model calls at some steps. - **Guide:** A feedforward harness control that steers the agent before it acts, such as AGENTS.md or a skill. - **Sensor:** A feedback harness control that checks the agent's work and returns the result so it can self-correct. - **Computational control:** A deterministic guide or sensor executed by code: tests, linters, type checkers, blocking hooks. - **Inferential control:** A guide or sensor where a model does the judging, such as a reviewer agent or LLM judge. - **Goal loop:** Repeating agent runs until a verifiable check passes or a limit is reached. - **Scheduled loop:** An agent run triggered repeatedly by a timer or event. - **Verified exit:** A loop stop condition checked by something other than the agent itself. - **No-progress check:** Stopping a loop when its progress measure stops improving for a set number of rounds. - **Structured notes:** Progress and decisions the agent writes to a file outside the context window and rereads later. ## k - **AGENTS.md:** An open Markdown convention for project instructions read by many coding agents; released by OpenAI in 2025 and stewarded by the Agentic AI Foundation. - **Bypass permissions:** A Claude Code permission mode that skips approval prompts; use only inside a sandbox. - **Checkpoint:** A saved state of the session and files that you can rewind to. - **CLAUDE.md:** Project instructions Claude Code loads at session start. - **Claude Code:** Anthropic's agentic coding tool for the terminal, IDE, desktop and web. - **Codex:** OpenAI's coding agent, available as a CLI and in the cloud. - **Exit code 2:** In a Claude Code hook, a blocking error: PreToolUse blocks the call and stderr goes to Claude. - **Headless mode:** Running Claude Code non-interactively, for example claude -p, in CI or scripts. - **Hook:** A command Claude Code runs on a lifecycle event; it can block, inject context or log. - **Permission modes:** Settings for what the agent may do without asking: read only, edit, run commands, bypass. - **Plan mode:** Claude Code mode where the agent proposes a plan before editing. - **Plugin:** A bundle of skills, hooks, commands, agents and MCP servers. - **PostToolUse:** Hook event after a tool runs; used for formatting, linting, logging and feedback. - **PreCompact:** Hook event that runs before compaction; used to save what matters. - **PreToolUse:** Hook event before a tool runs; can block, allow or modify the call. - **SessionStart:** Hook event at session start, resume or after compaction. - **Skill:** A folder with SKILL.md and scripts; its description stays loaded, its body loads on demand. - **Slash command:** A saved prompt triggered by name, such as /review. - **Statusline:** A command whose output Claude Code shows as a status bar: model, context, cost, session. - **Stop hook:** Hook event when the agent wants to finish; exit 2 forces it to continue. - **stop_hook_active:** Field telling a Stop hook that Claude is already continuing because of a Stop hook. - **Subagent:** A separate agent with its own context, started by a parent with a brief. - **SubagentStop:** Hook event when a subagent finishes. - **UserPromptSubmit:** Hook event when the user submits a prompt; can add context or block. - **Worktree:** A git feature for multiple checkouts of one repo; one per parallel agent. - **Agent Skills:** An open standard for packaging procedures as folders with a SKILL.md file plus optional scripts and references; specified at agentskills.io. - **goose:** An open source local agent framework from Block, stewarded by the Agentic AI Foundation. ## g - **A2A:** Agent2Agent protocol for agents from different systems to communicate and delegate. - **Backoff:** Waiting longer between retries, usually doubling each time, to recover from rate limits. - **Batch API:** Asynchronous bulk requests at lower cost, for work that is not urgent. - **Cost per accepted task:** Total spend divided by the number of runs whose output was accepted. - **Golden set:** A fixed set of real inputs with known good outputs, used for evals. - **HTTP+SSE transport:** MCP's original remote transport, replaced by Streamable HTTP and now deprecated. - **Latency:** Time to first token and total time to complete a response or run. - **LLM-as-judge:** Using a model to grade outputs against a rubric. - **MCP:** Model Context Protocol: an open standard that exposes tools, resources and prompts to agents. - **MCP prompts:** Reusable prompt templates an MCP server offers to users. - **MCP resources:** Read-only data an MCP server exposes to the model. - **MCP tools:** Actions an MCP server lets the model call. - **Observability:** Metrics, logs and traces across agent runs: cost, success rate, latency. - **OpenTelemetry:** An open standard for traces and metrics; MCP documents trace context propagation with it. - **Prompt caching:** Provider reuse of an unchanged prompt prefix, making repeat calls cheaper and faster. - **Rate limit:** A provider's cap on requests or tokens per minute, returned as HTTP 429. - **Regression eval:** Rerunning the eval suite after every change to catch quality drops. - **Sandbox:** An isolated environment, such as a container, VM or worktree, for running agents with wide permissions. - **Secrets handling:** Keeping keys out of prompts, context, logs and commits. - **stdio transport:** MCP transport where the server runs as a local process speaking over standard input and output. - **Streamable HTTP:** MCP's standard remote transport. - **Supply chain risk:** Risk from third-party plugins, skills and MCP servers that run with your access. - **tmux:** Terminal multiplexer with windows and panes that survive closing the terminal. - **Tool annotations:** MCP hints describing tool behaviour, such as read-only or destructive. - **Tracing:** Recording every step of a run, including tool calls, tokens, latency and cost. - **Vector store:** A database for embeddings and similarity search, such as Qdrant or pgvector. - **Agentic AI Foundation:** A Linux Foundation body launched in December 2025 to steward MCP, AGENTS.md and goose as open standards. - **pass@k:** Probability that at least one of k trials succeeds; a capability measure. - **pass^k:** Probability that all k trials succeed; a consistency measure. - **Trial:** One attempt at an eval task; agents are run several times per task because results vary. - **Grader:** Logic that scores an agent's output or transcript: code-based, model-based or human. - **Transcript:** The full record of an agent run: messages, tool calls and results.